Low-income families face increased exposure to stressors,1 2 such as material hardship, dispossession, limited social support,3 4 trauma, and violence,1 5 which subsequently affect family relationships and the physical and mental health of parents,6 7 8 contributing to household-wide feelings of stigma, isolation, and exclusion. These stressors are particularly relevant to Hong Kong, where approximately one-fifth of the population lives below the poverty line.9 Adults from low-income families in Hong Kong have reported significantly lower health-related quality of life (HRQOL) than age- and sex-matched individuals from the general population; low income is significantly associated with poorer mental health.10

Stressors may persist across the life course and affect the next generation, resulting in intergenerational socio-economic inequality and health disparities. Early caregiving experiences have been linked to later-life child health outcomes through physiological stress responses.11 Moreover, poor mental health in parents may lead to family disharmony and maladaptive parenting practices, which can increase a child’s risk of adverse health outcomes.7 8 Specifically, children of parents with depression tend to exhibit more difficult temperaments and diminished psychosocial functioning.12 13 Children from low-income families in Hong Kong have reported poorer health and more behavioural problems relative to population norms for similar age-groups.14 15 Without adequate parental care and guidance, such children may be more vulnerable to academic difficulties and behavioural problems, thereby exacerbating parental stress. A bidirectional relationship between parental stress and child health has been documented in Western studies6 8 but not within the Chinese context.

Stress coping can be mediated or moderated by various social factors.16 17 For instance, stressed parents may contribute to family disharmony, which mediates diminished child health. Neighbourhood cohesion may moderate this relationship by alleviating parental stress and enhancing children’s well-being. The identification of mediators and moderators that may influence the relationship between parental stress and child health enables development of targeted interventions and policy recommendations. Despite strong associations of parental depression with stress18 and child health,12 13 its mediating role in this relationship remains unclear. A recent study demonstrated mediation between parental stress and parent-infant bonding,19 but evidence concerning overall child health is lacking. This study aimed to explore whether a bidirectional relationship exists between parental stress and child health and to identify its mediators and moderators, with the goal of promoting health among parents and children from low-income families in Hong Kong. We hypothesised that parental stress precedes and results from child health, with mediating and moderating effects exerted by factors illustrated in Figure 1.

This prospective cohort study involved 217 parent-child pairs in which at least one parent was the primary caregiver and at least one parent was employed, with a monthly household income lower than 75% of the Hong Kong median at baseline. This income criterion included working poor families who lived above the poverty line (50% of the population median) and received limited government support. Families were recruited by research staff when attending health assessments during our previous cohort study20 performed in two less affluent Hong Kong communities between May 2016 and October 2017. Parents unable to communicate in Chinese, as well as children born prematurely and/or with congenital deformities, were excluded. All parents provided written informed consent for themselves and their child to participate in the study. Sample size was determined based on the need to detect a difference in Child Health Questionnaire (CHQ) scores between children of parents with high and low stress levels, classified according to the Depression Anxiety Stress Scales (DASS) stress subscale scores. Our previous cohort study showed that average CHQ general health perceptions subscale scores in children of parents with high and low DASS stress subscale scores were 59 (standard deviation [SD]=17) and 65 (SD=16), respectively20 (effect size=0.4). A sample size of 200 (100 per group) parent-child pairs was required to detect a difference of 6 points in CHQ general health perceptions subscale score between groups using an independent t test with 80% power and a 5% level of significance.

Each parent-child pair was invited to complete a comprehensive questionnaire survey at three time points (ie, baseline, 12 months, and 24 months) covering parental stress, HRQOL, and mental health; child’s general health, HRQOL, and behaviour; family harmony; parenting style; and neighbourhood cohesion, as reported by the parent. Potential confounders were recorded at baseline, including parental age, gender, education level, marital status, employment status, household income, smoking habits, and alcohol consumption, as well as the child’s age, gender, estimated intelligence quotient, and special education needs. Physical and mental co-morbidities in parents and children were recorded at all three time points.

Parental stress was measured using the stress subscale of the DASS–21 items questionnaire.21 A cut-off score of ≥15 indicated the presence of significant parental stress.21 The scale has been validated in a Chinese population.22

Child health was measured using the general health perceptions subscale score from the CHQ–Parent Form 50.23 A higher score indicates better perceived physical and psychological HRQOL in the child based on parental proxy report. The Chinese version has demonstrated good psychometric properties in local Chinese children.20

The Patient Health Questionnaire–9 (PHQ-9)24 was used to screen for parental depression. A cut-off score of ≥10 was regarded as clinically significant depression. The Chinese version of the PHQ-9 was validated and used in our previous study.20 Family harmony was measured using the Family Harmony Scale–Short Form (FHS-5).25 Higher single-factor harmony scores reflect greater harmony. The Chinese version has demonstrated good psychometric properties in local Chinese households.25 Parent-child interaction was assessed using the Child Physical Assault and Neglect subscales of the Parent–Child Conflict Tactics Scale (CTSPC).26 Higher scores indicate higher frequencies of respective issues in the past 12 months. The translated traditional Chinese version has demonstrated good psychometric properties.27 Parenting style was assessed using the Authoritative Parenting Style subscale of the short version of the Parenting Style and Dimensions Questionnaire.28 A higher score indicates a stronger tendency towards authoritative parenting. The questionnaire has been validated in the Chinese cultural setting.29 Neighbourhood support was measured using the Neighbourhood Collective Efficacy Scale.30 Higher scores indicate greater neighbourhood cohesion. The scale has been tested in Chinese in a local study.31

Baseline characteristics of parent-child pairs and their households were summarised using descriptive statistics. Differences between groups according to parental stress level were assessed using independent t tests for continuous variables and the Chi squared test for categorical variables.

The longitudinal bidirectional relationship between parental stress and child health was assessed using a cross-lagged panel model. Multiple indicators were utilised to evaluate model goodness-of-fit. A statistically non-significant Chi squared P value, Comparative Fit Index and Tucker-Lewis Index >0.95, root mean square error of approximation ≤0.05, and standardised root mean residual >0.08 were considered indicative of desirable goodness-of-fit. The final model was selected using root mean square error of approximation–based forward stepwise selection.

A mediation model was used to evaluate candidate mediators. Model estimates were obtained using 5000 bootstrapping samples. A statistically significant indirect effect, along with a reduced direct effect magnitude relative to the total effect, indicated that a given mediator explained the relationship between parental stress and child health.32 A multi-mediator model was constructed; differences in indirect effects between mediators were estimated via pairwise comparison.

Potential moderating effects of neighbourhood cohesion and parenting style on the relationship between parental stress and child health were examined by multivariable linear regression. A statistically significant interaction term coefficient indicated a moderation effect. All variables were centred to a mean of zero to reduce multicollinearity related to interaction terms. Confounders were included to improve model goodness-of-fit; R² and adjusted R² values were used to evaluate model performance.

All descriptive analyses were performed using Stata 16 (StataCorp LLC, College Station [TX], US); all model analyses were carried out using the lavaan package33 version 0.6-6, in R version 4.0.1 (R Foundation for Statistical Computing, Vienna, Austria). Data completion rates are presented in online supplementary Table 1. Complete case analyses were conducted. All tests were two-tailed; P values <0.05 were considered statistically significant.

Among the 217 parent-child pairs recruited at baseline, 175 (80.6%) and 184 (87.6%) pairs attended the 12-month and 24-month follow-ups, respectively (online supplementary Fig 1). Their characteristics at each of the three time points are detailed in Table 1.

At baseline, the ages of parents and children (mean ± SD) were 42.4 ± 6.2 years and 10.7 ± 2.0 years, respectively. Approximately half of the children were girls (47.5%), whereas the parents involved were predominantly mothers (91.7%). The majority (75.2%) of parents had completed secondary education. Approximately 39.8% of primary parents were employed, and 57.2% of families had a monthly household income below 75% of the 2016 Hong Kong median (ie, HK$25 000).34

Thirty-eight parents (17.5%) experienced significant stress, indicated by a DASS stress subscale score of 15 or above at baseline. Considerable differences were evident in their baseline characteristics compared with parents who were not stressed. Stressed parents were more likely to be single parents (41.2% vs 18.5%) and to have a household income below 50% of the Hong Kong median (50.0% vs 29.9%). A greater proportion of stressed parents reported being victims of intimate partner abuse (23.7% vs 10.9%). Diagnosed mental illnesses (23.7% vs 5.1%) and depression, indicated by a PHQ-9 score ≥10 (21.1% vs 2.4%), were more prevalent among these parents (Table 2). Both their physical and mental HRQOL were significantly worse (physical component score=42.5 ± 9.9 vs 49.1 ± 8.2; mental component score=38.1 ± 10.0 vs 55.5 ± 8.7; P<0.001).

Compared with children of parents who were not stressed, children of stressed parents were younger (age=10.0 ± 2.6 years vs 10.8 ± 1.8 years; P=0.020) and had worse general health and HRQOL, as reflected by lower scores in every subscale of the CHQ–Parent Form 50 except bodily pain and self-esteem. In particular, large differences were observed in four subscales: parental impact—emotional, parental impact—time, family activities, and family cohesion.

Moreover, stressed parents reported lower scores in family harmony (FHS-5) and neighbourhood cohesion (Neighbourhood Collective Efficacy Scale). Although parenting style did not differ significantly, stressed parents showed a greater tendency for physical punishment, as reflected by higher scores on the CTSPC–physical assault subscale, and for neglect, as indicated by higher CTSPC–neglect subscale scores, compared with parents who were not stressed (Table 2).

Figure 2 shows the cross-lagged panel model examining the bidirectional relationship between parental stress and child health. A bidirectional relationship between child health and parental stress was confirmed. Significant associations were observed between parental stress and child health at each time point (estimates: baseline=-0.22, 12 months=-0.21, 24 months=-0.47); between baseline child health and parental stress at 12 months (estimate=-0.40) and 24 months (estimate=-0.42); and between baseline parental stress and child health at 12 months (estimate=-0.57) and 24 months (estimate=-0.10).

The multi-mediation model results generated by bootstrapping are illustrated in Figure 3; the model estimates and goodness-of-fit statistics are presented in online supplementary Table 2. The total effect of the relationship between parental stress and child health was reduced when mediators were included in the model. Significant positive associations of parental stress were observed with the PHQ-9 score, as well as the physical assault and neglect subscales of the CTSPC. A significant negative association was noted between parental stress and the FHS-5 score. Among mediators, only the PHQ-9 exerted a significant negative effect on child health.

Table 3 presents the moderation model. Neither neighbourhood cohesion nor parenting style demonstrated a moderating effect on the relationship between parental stress and child health. Estimates for the interaction terms were negligible. The R² values were around 0.21, and the adjusted R² values were slightly lower (0.11-0.13), indicating modest explanatory power of the model after adjusting for confounders.

Our study demonstrated that a substantial proportion of low-income parents experienced stress (17.5%), which was associated with multiple stressors including poverty, marital problems, intimate partner abuse, family disharmony, and reduced neighbourhood support. Children of stressed parents reported worse general health and HRQOL, as well as more behavioural problems. A short-term and long-term bidirectional inverse relationship between parental stress and child health was confirmed; this relationship was partially mediated by the level of parental depression.

Compared with the general Hong Kong population, the parent-child pairs in this study were more exposed to various known stressors in addition to low income. The prevalences of single-parent families (22.3% vs 9.8%35) and intimate partner abuse (13.2% vs 7.2%36) were higher, and more parents reported regular alcohol consumption (17.4% vs 8.7%37). Therefore, it is not surprising that a considerably greater proportion of parents in this study experienced elevated levels of stress (17.5% vs 5.2%38) and depression (5.9% vs 1.2%37). The persistently high level of parental stress observed during the study period may be attributed partly to ongoing exposure to various stressors over time and partly to constant exposure to chronic stressors. Both scenarios highlight the urgent need to ensure assessment and intervention for these disadvantaged parents.

Previous studies have demonstrated bidirectional interactions between parental stress and child health in relation to both internalising and externalising behaviours.6 8 Increases in behavioural problems have been shown to raise parental stress over time, which in turn exacerbates behavioural issues in children.39 Our study adds to this body of evidence by confirming significant bidirectional effects between general parental stress and child health at each time point. Cross-effects were observed from baseline child health to later parental stress, and from baseline parental stress to later child health at both 12 and 24 months. These findings suggest that parental stress both precedes and results from child health, with reciprocal short-term and long-term influences.40

In our attempt to identify pathways through which parental stress affects child health, we observed that only parental depression significantly mediated the relationship. This result is consistent with previous findings that maternal depression and perceived stress directly and negatively influence child development.41 One possible explanation is that depressed mothers may lack the energy or capacity to provide adequate care and support for their child’s health. Research into this mediation effect remains limited; however, one recent study reported similar outcomes regarding the indirect impact of workrelated stress on child health, mediated by maternal depression.42 The implementation of screening and intervention for parental depression is both imperative and urgent to counteract the adverse effects of stress on parental and child health. Medical and social service providers should collaborate to actively screen at-risk parents from low-income families in the community. Early intervention through lifestyle-based care—such as physical activity, relaxation techniques, and mindfulness-based therapies—can help to prevent43 44 and manage45 46 depression, thus mitigating long-term negative impacts on child health.

However, it must be noted that parents with depression may be biased towards over-reporting their child’s problems,47 compared with other informants such as teachers and the children themselves.48 Further research is warranted to identify individual and family characteristics that may influence discrepancies between informants. Other potential factors examined in previous studies—such as household structure (dual- vs multi-generational), parental rearing behaviours, and confident and affective social support—might also contribute to the relationship between parental stress and child health; they should be explored in future studies with larger sample sizes.

This is one of the first studies to examine the longitudinal relationship between general parental stress and child health, enabling assessment of possible causal relationships between the two outcomes. Specifically, we recruited vulnerable families with substantial socio-economic disadvantages who experience high levels of stress and would benefit most from future interventions. Furthermore, a high response rate was maintained throughout the study, ensuring adequate power for the analyses.

However, the findings of our study must be interpreted in light of the following limitations. First, although we conducted a comprehensive analysis of factors related to parental stress and child health, the outcomes were based on self-reported assessments, which are susceptible to respondent bias. Only three measurements, taken 1 year apart, were performed in this study due to concerns regarding practicality and the burden on participating families. Therefore, caution should be exercised in generalising the results with respect to longitudinal trends, given that substantial intra-individual fluctuations may have occurred but were not captured in this study. Second, both parental stress and child health were assessed using parent-report questionnaires, which may contribute to increased shared method variance. Additionally, aspects of the child’s health or behaviour considered problematic by the parent may not align with assessments made by other individuals (eg, teachers). As mentioned earlier, parents with depression may be biased towards over-reporting problems and are more likely to report behavioural issues in their child compared with other informants.47 48 The validity of parent-perceived measures of child health—particularly in relation to parental depression—and their agreement with other caregivers should be examined in future trials specifically designed for this purpose. Third, there were unmeasured confounders in this observational study, such as exercise and social functioning. Moreover, certain socio-demographic factors, including marital and employment statuses, were assumed to be static throughout the study. It remains uncertain whether changes in these factors, if any, may have influenced the observed results. Additional information regarding participant characteristics, observational measures of child behaviour, or objective indicators of child health (eg, cortisol levels) could improve the reliability of the findings.

This study showed that a substantial proportion of parents from low-income families in Hong Kong experienced general stress due to multiple stressors, which was negatively associated with their child’s health. A bidirectional relationship was observed between parental stress and child health over time, which may be partly mediated by parental depression. Prompt screening and appropriate intervention are necessary to prevent adverse health outcomes for parents and children in low-income families.

Concept or design: EYT Yu, RSM Wong, AFY Tiwari, CKH Wong, VY Guo, CLK Lam.
Acquisition of data: RSM Wong, KSN Liu.
Analysis or interpretation of data: EYT Yu, EYF Wan, RSM Wong, IL Mak, AFY Tiwari, CKH Wong, VY Guo, CLK Lam.
Drafting of the manuscript: EYT Yu, RSM Wong, IL Mak, CHN Yeung.
Critical revision of the manuscript for important intellectual content: All authors.

All authors had full access to the data, contributed to the study, approved the final version for publication, and take responsibility for its accuracy and integrity.

As advisors of the journal, EYT Yu and CKH Wong were not involved in the peer review process. Other authors have disclosed no conflicts of interest.

The authors are grateful for the support from Kerry Group Kuok Foundation (Hong Kong) Limited in conducting this study on participants of the Trekkers Family Enhancement Scheme. The authors’ sincere gratitude goes to the Neighbourhood Advice-Action Council, Hong Kong Sheng Kung Hui Lady MacLehose Centre, and Shek Lei Community Hall for their assistance in participant recruitment and provision of venues for data collection, respectively. The authors thank the Social Science Research Centre of The University of Hong Kong (HKU) for their timely completion of the telephone surveys, and Department of Paediatrics and Adolescent Medicine of HKU for performing the assays for DNA extraction and telomere length measurement. The authors also thank the hard work of their research staff in data collection and analysis.

The study results were disseminated through a poster presentation at the Health Research Symposium 2021 (23 November 2021, hybrid conference), entitled “In-depth exploration of a bidirectional parent-child health relationship and its mediating and moderating factors among low-income families in Hong Kong”.

This research was supported by the Health and Medical Research Fund of the Health Bureau, Hong Kong SAR Government (Ref No.: HMRF 14151571). The funder had no role in the study design, data collection/analysis/interpretation, or manuscript preparation.

This research was approved by the Institutional Review Board of The University of Hong Kong/Hospital Authority Hong Kong West Cluster, Hong Kong (Ref No.: UW 16-415). Informed consent was obtained from patients when baseline data were collected.

The supplementary material was provided by the authors and some information may not have been peer reviewed. Accepted supplementary material will be published as submitted by the authors, without any editing or formatting. Any opinions or recommendations discussed are solely those of the author(s) and are not endorsed by the Hong Kong Academy of Medicine and the Hong Kong Medical Association. The Hong Kong Academy of Medicine and the Hong Kong Medical Association disclaim all liability and responsibility arising from any reliance placed on the content.

1. Santiago CD, Kaltman S, Miranda J. Poverty and mental health: how do low-income adults and children fare in psychotherapy? J Clin Psychol 2013;69:115-26. Crossref

2. Smith MV, Mazure CM. Mental health and wealth: depression, gender, poverty, and parenting. Annu Rev Clin Psychol 2021;17:181-205. Crossref

3. Evans GW, Kim P. Childhood poverty, chronic stress, self-regulation, and coping. Child Dev Perspect 2013;7:43-8. Crossref

4. Adjei NK, Jonsson KR, Straatmann VS, et al. Impact of poverty and adversity on perceived family support in adolescence: findings from the UK Millennium Cohort Study. Eur Child Adolesc Psychiatry 2024;33:3123-32. Crossref

5. Alto ME, Warmingham JM, Handley ED, Manly JT, Cicchetti D, Toth SL. The association between patterns of trauma exposure, family dysfunction, and psychopathology among adolescent females with depressive symptoms from low-income contexts. Child Maltreat 2023;28:130-40. Crossref

6. van Dijk W, de Moor MH, Oosterman M, Huizink AC, Matvienko-Sikar K. Longitudinal relations between parenting stress and child internalizing and externalizing behaviors: testing within-person changes, bidirectionality and mediating mechanisms. Front Behav Neurosci 2022;16:942363. Crossref

7. Neece CL, Green SA, Baker BL. Parenting stress and child behavior problems: a transactional relationship across time. Am J Intellect Dev Disabil 2012;117:48-66. Crossref

8. Stone LL, Mares SH, Otten R, Engels RC, Janssens JM. The co-development of parenting stress and childhood internalizing and externalizing problems. J Psychopathol Behav Assess 2016;38:76-86. Crossref

9. Economic Analysis Division Economic Analysis and Business Facilitation Unit Financial Secretary’s Office; Census and Statistics Department, Hong Kong SAR Government. Hong Kong Poverty Situation Report 2013. Oct 2014. Available from: https://www.commissiononpoverty.gov.hk/eng/pdf/poverty_report13_rev2.pdf. Accessed 31 Jul 2023.

10. Lam CL, Guo VY, Wong CK, Yu EY, Fung CS. Poverty and health-related quality of life of people living in Hong Kong: comparison of individuals from low-income families and the general population. J Public Health (Oxf) 2017;39:258-65.Crossref

11. Luecken LJ, Lemery KS. Early caregiving and physiological stress responses. Clin Psychol Rev 2004;24:171-91. Crossref

12. Hanington L, Ramchandani P, Stein A. Parental depression and child temperament: assessing child to parent effects in a longitudinal population study. Infant Behav Dev 2010;33:88-95. Crossref

13. Associations between depression in parents and parenting, child health, and child psychological functioning. In: England MJ, Sim LJ, editors. Depression in Parents, Parenting, and Children: Opportunities to Improve Identification, Treatment, and Prevention. Washington (DC): National Academies Press (US); 2009: 119-82.

14. Lee SL, Cheung YF, Wong HS, Leung TH, Lam T, Lau YL. Chronic health problems and health-related quality of life in Chinese children and adolescents: a population-based study in Hong Kong. BMJ Open 2013;3:e001183. Crossref

15. Chan KL, Lo CK, Ho FK, Chen Q, Chen M, Ip P. Modifiable factors for the trajectory of health-related quality of life among youth growing up in poverty: a prospective cohort study. Int J Environ Res Public Health 2021;18:9221. Crossref

16. Asok A, Bernard K, Roth TL, Rosen JB, Dozier M. Parental responsiveness moderates the association between early-life stress and reduced telomere length. Dev Psychopathol 2013;25:577-85. Crossref

17. Evans GW, Kim P, Ting AH, Tesher HB, Shannis D. Cumulative risk, maternal responsiveness, and allostatic load among young adolescents. Dev Psychol 2007;43:341-51. Crossref

18. Hammen C. Stress and depression. Annu Rev Clin Psychol 2005;1:293-319. Crossref

19. Power C, Weise V, Mack JT, Karl M, Garthus-Niegel S. Does parental mental health mediate the association between parents’ perceived stress and parent-infant bonding during the early COVID-19 pandemic? Early Hum Dev 2024;189:105931. Crossref

20. Fung CS, Yu EY, Guo VY, et al. Development of a Health Empowerment Programme to improve the health of working poor families: protocol for a prospective cohort study in Hong Kong. BMJ Open 2016;6:e010015. Crossref

21. Lovibond SH, Lovibond PF; Psychology Foundation of Australia. Manual for the Depression Anxiety Stress Scales. Sydney: Sydney Psychology Foundation; 1995. Crossref

22. Wang K, Shi HS, Geng FL, et al. Cross-cultural validation of the Depression Anxiety Stress Scale–21 in China. Psychol Assess 2016;28:e88-100. Crossref

23. Landgraf JM. Child Health Questionnaire (CHQ). In: Maggino F, editor. Encyclopedia of Quality of Life and Well-being Research. Cham: Springer; 2020: 1-6. Crossref

24. Kroenke K, Spitzer RL, Williams JB. The PHQ-9: validity of a brief depression severity measure. J Gen Intern Med 2001;16:606-13. Crossref

25. Kavikondala S, Stewart SM, Ni MY, et al. Structure and validity of Family Harmony Scale: an instrument for measuring harmony. Psychol Assess 2016;28:307-18. Crossref

26. Straus MA. Measuring intrafamily conflict and violence: the Conflict Tactics (CT) Scales. J Marriage Fam 1979;41:75-88. Crossref

27. Chan KL, Brownridge DA, Fong DY, Tiwari A, Leung WC, Ho PC. Violence against pregnant women can increase the risk of child abuse: a longitudinal study. Child Abuse Negl 2012;36:275-84. Crossref

28. Robinson CC, Mandleco B, Olsen SF, Hart CH. The Parenting Styles and Dimensions Questionnaire (PSDQ). In: Perlmutter BF, Touliatos J, Holden GW, editors. Handbook of Family Measurement Techniques: Vol 3. Instruments & Index. Thousand Oaks: Sage; 2001: 319-21.

29. Wu P, Robinson CC, Yang C, et al. Similarities and differences in mothers’ parenting of preschoolers in China and the United States. Int J Behav Dev 2002;26:481-91. Crossref

30. Sampson RJ, Raudenbush SW, Earls F. Neighborhoods and violent crime: a multilevel study of collective efficacy. Science 1997;277:918-24. Crossref

31. Chou KL. Perceived discrimination and depression among new migrants to Hong Kong: the moderating role of social support and neighborhood collective efficacy. J Affect Disord 2012;138:63-70. Crossref

32. Baron RM, Kenny DA. The moderator–mediator variable distinction in social psychological research: conceptual, strategic, and statistical considerations. J Pers Soc Psychol 1986;51:1173-82. Crossref

33. Rosseel Y. lavaan: an R package for structural equation modeling. J Stat Softw 2012;48:1-36. Crossref

34. Census and Statistics Department, Hong Kong SAR Government. Hong Kong 2016 Population By-census–Thematic Report: Household Income Distribution in Hong Kong. Jun 2017. Available from: https://www.censtatd.gov.hk/en/data/stat_report/product/B1120096/att/B11200962016XXXXB0100.pdf. Accessed 25 Aug 2025.

35. Census and Statistics Department, Hong Kong SAR Government. 2021 Population Census—Thematic Report: Children. Feb 2023. Available from: https://www.census2021.gov.hk/doc/pub/21c-Children.pdf. Accessed 25 Aug 2025.

36. Chan KL. Intimate partner violence in Hong Kong. In: Chan KL, editor. Preventing Family Violence: A Multidisciplinary Approach. Hong Kong: Hong Kong University Press; 2012: 19-58. Crossref

37. Non-Communicable Disease Branch, Centre for Health Protection, Hong Kong SAR Government. Report of Population Health Survey 2020-22 (Part I); 2022. Available from: https://www.chp.gov.hk/files/pdf/dh_phs_2020-22_part_1_report_eng_rectified.pdf. Accessed 31 Jul 2023.

38. Chan SM, Wong H, Chung RY, Au-Yeung TC. Association of living density with anxiety and stress: a cross-sectional population study in Hong Kong. Health Soc Care Community 2021;29:1019-29. Crossref

39. Baker BL, McIntyre LL, Blacher J, Crnic K, Edelbrock C, Low C. Pre-school children with and without developmental delay: behaviour problems and parenting stress over time. J Intellect Disabil Res 2003;47:217-30. Crossref

40. Motrico E, Bina R, Kassianos AP, et al. Effectiveness of interventions to prevent perinatal depression: an umbrella review of systematic reviews and meta-analysis. Gen Hosp Psychiatry 2023;82:47-61. Crossref

41. Vameghi R, Amir Ali Akbari S, Sajedi F, Sajjadi H, Alavi Majd H. Path analysis association between domestic violence, anxiety, depression and perceived stress in mothers and children’s development. Iran J Child Neurol 2016;10:36-48. Crossref

42. Xu L, Xu J. The impact of maternal occupation on children’s health: a mediation analysis using the parametric G-formula. Soc Sci Med 2024;343:116602. Crossref

43. Bellón JÁ, Conejo-Cerón S, Sánchez-Calderón A, et al. Effectiveness of exercise-based interventions in reducing depressive symptoms in people without clinical depression: systematic review and meta-analysis of randomised controlled trials. Br J Psychiatry 2021;219:578-87. Crossref

44. Newland P, Bettencourt BA. Effectiveness of mindfulness-based art therapy for symptoms of anxiety, depression, and fatigue: a systematic review and meta-analysis. Complement Ther Clin Pract 2020;41:101246. Crossref

45. Marx W, Manger SH, Blencowe M, et al. Clinical guidelines for the use of lifestyle-based mental health care in major depressive disorder: World Federation of Societies for Biological Psychiatry (WFSBP) and Australasian Society of Lifestyle Medicine (ASLM) taskforce. World J Biol Psychiatry 2023;24:333-86. Crossref

46. Recchia F, Leung CK, Chin EC, et al. Comparative effectiveness of exercise, antidepressants and their combination in treating non-severe depression: a systematic review and network meta-analysis of randomised controlled trials. Br J Sports Med 2022;56:1375-80. Crossref

47. Chi TC, Hinshaw SP. Mother–child relationships of children with ADHD: the role of maternal depressive symptoms and depression-related distortions. J Abnor Child Psychol 2002;30:387-400. Crossref

48. Richters JE. Depressed mothers as informants about their children: a critical review of the evidence for distortion. Psychol Bull 1992;112:485-99. Crossref

Clinical research is fundamental to the advancement of medicine. More than a quarter of a century on, evidence-based medicine—which began as a nascent movement in the early 1990s—has revolutionised healthcare by producing trustworthy observations that support better-informed clinical decision-making and health policy.1 2 Research forms the foundation of evidence-based medicine and plays an important role in understanding disease, thereby contributing to the development of novel therapeutic strategies.3 This contribution has translated into quantifiable outcomes: participation in clinical research can lead to significant reductions in patient mortality and inpatient length of stay (LOS).4 5 6 7 8 9 Clinical research benefits individual patients and drives socio-economic growth. The UK National Institute for Health and Care Research (NIHR) observed that every 1.0 GBP invested in clinical trials yielded a return of up to 7.6 GBP in economic benefit.10 However, a substantial proportion of frontline clinicians typically do not engage in research activities relevant to their daily practice. A cross-sectional survey in North America revealed that 32% of respondents did not know how to participate in research.11 A similar study among Hong Kong family physicians indicated that 27% had no previous experience.3 Hong Kong is an ideal location for conducting clinical research due to its world-class universal healthcare infrastructure, electronic medical records system, use of English in medical documentation, and the presence of a pool of internationally reputable investigators.12 13 Additionally, the Hospital Authority (HA)—a statutory body responsible for managing all public hospitals in the city—provides more than 90% of all inpatient bed-days, and the patient follow-up rate is comparably high.14 Regardless of these favourable factors, according to the Our Hong Kong Foundation—a non-governmental, non-profit public policy institute—the number of clinical trials conducted in Hong Kong declined by 22% between 2015 and 2021, compared with a mean increase of 48% in developed countries and 285% in Mainland China.15 No comprehensive review of the clinical research activity of Hong Kong public healthcare professionals has been conducted. Apart from the UK and Spain, no other region has evaluated the influence of clinician engagement in research on key performance indicators within a universal healthcare system.4 5 6 This study was performed to determine Hong Kong’s research productivity in terms of peer-reviewed published clinical studies, its scholarly impact, and its influence on outcomes for hospitalised patients—including LOS, crude mortality, and 30-day unplanned readmission. A comparative analysis of research productivity and quality across medical disciplines was also performed. Findings from this review could inform health policy by providing a stronger foundation for the evidence-based allocation of resources to support an efficient and sustainable research ecosystem within the HA.

This was a territory-wide retrospective observational study of peer-reviewed original clinical research conducted by HA medical staff at general acute care hospitals, in which the staff member served as principal investigator. The review included articles published in the biomedical literature from 1 January 2016 to 30 April 2021. Research articles from non-university institutions—comprising 30% (13/43) of all HA hospitals—were included. Citations covering this 5.5-year period were retrieved from MEDLINE, the United States National Library of Medicine’s bibliographic database. The database was queried via the PubMed Advanced Search Builder for all studies published within the review period, where the first author’s stated affiliation was a Hong Kong hospital. The internet-based library search package RISmed was used to extract author affiliation data from the PubMed search results into R, an open-source statistical software tool.16 17 The list of citations was then manually reviewed to confirm that the study had been conducted by a clinician from an HA hospital. Published abstracts were categorised according to study design, article type, and corresponding medical specialty (Table 1).18 Systematic reviews performed in accordance with the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) 2020 statement were regarded as original research.19 20 Articles not subject to peer review—such as academic conference proceedings, trial protocols, editorials, letters to the editor, and erratum or corrigendum statements—were excluded. Preclinical studies and secondary research articles, including clinical practice guidelines, position statements, book chapters, and narrative topical reviews, were also excluded. Finally, collaborative studies in which the principal investigator was not employed by the HA were excluded.

The primary study endpoint was research productivity, measured by the total number of original research studies published, with comparisons made between university- and non–university-affiliated HA hospitals. The hypothesis was that university-affiliated hospitals would produce more original clinical research studies than non-university hospitals because of their access to tertiary education institution resources. Secondary endpoints included research productivity across medical specialties. To control for workforce discrepancies across medical disciplines, the mean number of full-time clinicians for each specialty from 2016 to 2021 was determined. The number of articles per clinician and the proportion of the workforce acting as principal investigator for each specialty were then established. The quality of the research, as reflected by the scientometric performance of each published article, was also assessed. It was hypothesised that research quality from university-affiliated hospitals would be superior to that of their non-university counterparts. Another secondary endpoint was patient outcomes for each non–university-affiliated acute care hospital from 2016 to 2021: crude mortality rate per 100 000 hospitalised patients, length of inpatient stay, and annual number of unplanned readmissions within 30 days of discharge. It was hypothesised that increased research productivity would translate to improved patient outcomes, and an inter-hospital comparison of these key performance indicators was performed. Patient outcome data were collected from the HA’s Clinical Data Analysis and Reporting System and the HA Management Information System.

To evaluate research quality, a multi-level scientometric approach was utilised by collecting journal-, article-, and individual author-level data. For journal-level scientometric assessment, two indices were determined: the journal’s mean impact factor (IF) and Eigenfactor score (ES) from 2016 to 2021. These indices were obtained from Clarivate (London, UK), a bibliometric analytics company that manages the Science Citation Index, an online indexing database containing academic journal citation data.1 A journal’s IF is a scientometric index reflecting the mean number of citations received per article in that journal during the preceding 2 years.21 This metric constitutes a reasonable indicator of research quality for general medical journals.22 The ES ranks journals using eigenvector centrality statistics to evaluate the importance of citations within a scholarly network.23 Utilising an algorithm similar to Google’s PageRank (Alphabet Inc, Mountain View [CA], US), the ES considers the number of citations received and the prestige of the citing journal. For article-level metrics, the total number of citations per article (TNC), Relative Citation Ratio (RCR), Field Citation Ratio (FCR), and National Institutes of Health (NIH) percentile attained were documented (Table 2). Author-level data were collected by determining the h-index of the principal investigator (Table 2).24 All scientometric data were censored on 30 June 2023.

Independent-samples t tests and Chi squared tests were conducted to compare variables. Spearman’s rank analysis was performed to assess correlations between research and hospitalised patient outcomes. P values of less than 0.05 were considered statistically significant. All statistical tests were performed using SPSS (Windows version 21.0; IBM Corp, Armonk [NY], US) and R (version 4.5.0; R Foundation, Vienna, Austria).17

During the 5.5-year period, 4511 peer-reviewed articles were published by Hong Kong medical researchers from acute care public hospitals. Of these, 3142 (69.7%) were original clinical research studies. In total, 29.3% (n=921) of the articles were authored by non-university hospital investigators—a significantly smaller proportion than that published by their university hospital counterparts (independent-samples t test, P<0.001) [Fig 1a]. Throughout the review period, the annual number of publications by non-university hospital investigators remained consistent, with a mean of 167 ± 8 per year (t test, P=0.24) [Fig 1b]. Overall, the medical specialties that produced the most articles were internal medicine, representing 23.4% (n=735) of published studies, and general surgery, representing 16.5% (n=520) [Fig 2].

The majority of excluded articles were narrative reviews that did not meet PRISMA criteria, followed by letters to the editor and editorials (Fig 1a). Regarding study design, most research articles were case reports, followed by retrospective cohort and prospective cohort studies (Fig 3). A significantly larger proportion of case reports were published by non-university hospital staff compared with university hospital staff (P<0.001). In addition to retrospective studies, university hospital investigators were significantly more likely to publish higher level-of-evidence research articles (Table 3).

The majority of non-university hospital principal investigators were clinicians, with 3.7% (n=34) of studies conducted by nurses or allied healthcare professionals. Among the 887 articles authored by clinicians, 544 individuals were identified, yielding an author-to-article ratio of 1:1.6. These researchers comprised 10.8% of the 5056 full-time non-university hospital clinicians employed during the study period. The most research-productive specialties among non-university hospitals were orthopaedics, followed by internal medicine and obstetrics and gynaecology (Table 4). Among all the medical specialties, the mean number of articles per clinician (×100) was 17.5 ± 22.3 (range, 1.8-56.4). After controlling for workforce discrepancies between disciplines, clinical oncology, orthopaedics and traumatology, and obstetrics and gynaecology constituted the most productive specialties in terms of the mean number of articles published per clinician (Table 4). Collectively, these three specialties published significantly more studies than the other disciplines (independent-samples t test, P<0.001) [Table 4]. The most research-active specialty was obstetrics and gynaecology, where one-third of clinicians acted as principal investigators—this proportion was significantly larger relative to other medical disciplines (t test, P=0.03) [Table 4].

Regarding scientometric performance, the overall mean journal IF and ES were 2.34 ± 3.72 and 0.01 ± 0.07, respectively. No statistically significant difference was observed between the principal investigator’s medical specialty and the journal IF (independent-samples t test, P=0.31). However, with respect to the ES, clinical oncologists published their research in journals with a significantly higher score relative to other medical disciplines (P<0.001) [Table 5].

For studies performed by non-university clinicians during this period, at the individual article level, the mean TNC per study was 6.33 ± 24.17, the mean number of citations per year was 1.81 ± 9.52, mean RCR was 0.82 ± 3.32, the mean FCR was 3.37 ± 2.04, and the mean NIH percentile achieved was 28.73 ± 25.85. Combined, radiology and otorhinolaryngology research articles had a significantly higher TNC per study (P<0.01) and a higher total number of annual citations (P<0.001) compared with other medical disciplines (Table 5). Articles in anaesthesiology, ophthalmology, otorhinolaryngology, and radiology had a mean RCR exceeding 1.00, indicating that their articles received a higher citation rate than their co-citation network. In particular, anaesthesiology and ophthalmology, studies achieved the highest mean NIH percentile rankings: their research outperformed 47% of all NIH-associated publications. Combined, anaesthesiology and ophthalmology studies also had a significantly higher NIH percentile ranking than other medical disciplines (P=0.001). All medical specialties had an FCR exceeding 1.00, indicating that their article citation rates were higher than those of their counterparts in the same research field. Oncology research had a significantly higher mean FCR (5.82 ± 2.41) compared with other disciplines (P<0.001) [Table 5].

In terms of author-level scientometric performance, 18.0% (98/544) of authors did not have a documented h-index. The mean h-index for the remaining researchers was 7.54 ± 10.98. Anaesthesiologists had a significantly higher h-index relative to other specialties (independent-samples t test, P=0.01) [Table 5]. A comparison of scientometric outcomes between university and non-university clinical research also demonstrated uniformly superior performance by academic institution investigators (Table 6).

For non-university-affiliated hospitals, the overall mean crude mortality rate per 100 000 hospitalised patients was 27 722 ± 5208. Spearman’s rank analysis identified significant negative correlations of mortality rate with TNC (r=-0.69; P=0.01) and RCR (r=-0.63; P=0.022). The overall annual mean number of unplanned readmissions within 30 days of discharge was 1408 ± 756. Similarly, there were significant negative correlations of readmissions with TNC (r=-0.76; P=0.02) and RCR (r=-0.72; P=0.006). The overall mean LOS was 11.7 ± 3.1 days. No significant correlations between LOS and TNC (r=-0.32; P=0.29) or LOS and RCR (r=-0.36; P=0.23) were detected. None of the other scientometric indices were associated with the crude mortality rate, number of unplanned readmissions, or LOS.

This study reviewed the breadth and quality of original clinical research conducted by Hong Kong’s public healthcare professionals. It is encouraging to observe that, despite the heavy workload of frontline clinicians employed in non-university public hospitals, more than 10% of the workforce engaged in original research as principal investigators. Their endeavours contributed to nearly one-third of peer-reviewed publications produced in the territory. A multi-level scientometric approach was adopted to assess research quality, and our findings indicate that the studies undertaken met the standards of their respective fields. Although the IF and ES values of the published research were not high, all medical specialties achieved a mean FCR of over 1.00. Notably, anaesthesiology, ophthalmology, otorhinolaryngology, and radiology articles attained an RCR exceeding 1.00.

Introduced in 2016, the RCR is a relatively novel article-level metric that measures a publication’s relevance within the biomedical literature.24 It was developed in response to the limitations of conventional indicators of scientific quality, such as the IF25 and h-index.26 For example, as multidisciplinary collaborations have become more common, researchers in disparate fields may have unequal access to high-profile journals, undermining the IF as a reliable reflection of a study’s performance.21 24 25 Conversely, the h-index does not consider an author’s total number of citations and instead reflects cumulative output, which can disadvantage early-career researchers. Despite their limitations, the IF25 and h-index26 remain pivotal scientometric indices in decisions related to funding and career progression. Given that citations are widely recognised as a form of acknowledging a researcher’s contribution to the field, efforts have been made to formalise this practice into a quantifiable metric. Endorsed by the NIH, the RCR harnesses an article’s co-citation network, normalising the number of citations received according to the article’s publication time and field of expertise. It is calculated as the ratio of the article’s actual citation rate—derived from the FCR—to the expected rate, benchmarked against NIH-funded publications issued in the same year and specialty.24 In recent years, the RCR has gained recognition as a more reliable indicator of an article’s performance within its peer comparison group and is increasingly cited in research grant applications.27 28 29

The present study showed that university hospitals not only outperformed non-university hospitals in terms of research productivity, but also demonstrated greater influence across all scientometric outcomes. In addition to resource consolidation and the employment of clinician-scientists, another reason for this discrepancy might be the type of studies produced. Approximately half of the articles from non-university hospitals were case reports or technical notes, which provide a lower level of evidence in the evidence-based medicine hierarchy and consequently tend to receive fewer citations. Nonetheless, this form of research is more accessible to junior clinicians and can serve as a gateway to medical writing in resource-limited settings.30 Case reports offer valuable insights into the real-world implications of clinical practice—findings that well-designed randomised controlled trials may fail to capture. They can also stimulate others to report similar observations, serving as a hypothesis-generating opportunity for subsequent systematic enquiry.31

Few studies have tested the hypothesis that research activity results in improved patient outcomes.4 5 7 8 32 We observed that non-university hospitals whose staff engaged in clinical research had lower crude mortality rates and annual 30-day unplanned readmissions. These findings are supported by reports that patients treated at hospitals participating in clinical trials fared better in terms of 30-day post-intervention mortality and overall survival, relative to those treated at hospitals without such arrangements. This trend has been observed for conditions including acute myocardial infarction, small-cell lung cancer, colorectal cancer, breast cancer, and ovarian cancer.5 33 34 35 36 37 The possibility of a trial effect was reinforced by a systematic review of 13 studies, which attributed this phenomenon to healthcare providers’ greater adherence to clinical practice guidelines and their inclination to adopt evidence-based practices.8 A subsequent systematic review of 33 studies further demonstrated that research activity improved healthcare system performance—reflected by reductions in LOS and risk-adjusted mortality, as well as improvements in patient satisfaction.9 In contrast, few studies have quantitatively analysed peer-reviewed scientometric data and its relationship with patient outcomes. For specific disease conditions, a negative correlation was observed between acute myocardial infarction–related risk-adjusted mortality and a weighted citations ratio among 50 Spanish hospitals.7 A review of 147 National Health Service trusts in the UK demonstrated a negative correlation between the number of research article citations per admission and standardised mortality ratios.5 Econometric modelling using data from 189 Spanish hospitals detected a significant reduction in LOS among institutions that published more clinical research articles or had a higher TNC per article.6

There is increasing evidence that clinical research engagement improves patient outcomes, but several barriers to participation remain. First, clinicians have demanding responsibilities that often prohibit involvement in this time-consuming and resource-intensive activity.38 39 40 Clinicians are under-recognised for their overtime efforts—when such work is typically undertaken—and are overburdened with administrative procedures. Research-supportive policies that provide protected time or incentivise clinicians through career advancement could help foster a more scholastic environment.40 Second, Hong Kong has a lengthy and duplicative clinical trial approval process. In a survey of 250 clinician-researchers, 90% reported that approval for a phase I first-in-human study certificate from the Hong Kong SAR Government’s Department of Health required over 3 months.15 Additionally, for HA Clinical Research Ethics Committee study approvals, 50% of respondents reported that the process typically lasted more than 3 months, whereas multi-centre trials frequently required over a year to begin recruitment.15 The establishment of a primary review authority for investigative drug registration—similar to the United States Food and Drug Administration, European Medicines Agency, or China’s National Medical Products Administration—could help streamline regulatory pathways. Third, most funding agencies favour academician-led research over community clinician-led efforts.38 For example, the existing Hong Kong SAR Government’s Health and Medical Research Fund and the Health Care and Promotion Fund—with a combined annual budget of US$530 million—have primarily been allocated to academicians with access to robust research infrastructure. The lack of financial support for community hospitals to develop research capabilities can have clinical implications.4 38 39 40 A review of funding allocations from the NIHR revealed that National Health Service trusts receiving relatively lower levels of research funding had higher risk-adjusted mortality.4 A survey of healthcare professionals in Ontario, Canada, showed that 46% were dissatisfied with their research involvement, although 83% agreed it benefited their careers.39 The major barriers identified were a lack of mentorship and institutional stewardship.39 The establishment of a clinical research institute and academy dedicated to supporting early-career clinician-scientists could help address these challenges.15 Modelled after the NIHR, the provision of publicly funded administrative services to accelerate translational research—by facilitating grant applications for non–university-affiliated hospitals, offering biostatistical support, training research support staff, and nurturing partnerships in a multi-stakeholder ecosystem—can be transformative.40 Following the introduction of NIHR services, there was a tenfold rise in publications, accompanied by a significant increase in mean citation ratios.41 A survey of NIHR stakeholders—including clinicians, nurses, and allied health professionals—also revealed that its training programmes enhanced their research capacity and strengthened individual career development.42

This study had several limitations. First, we retrieved studies only from the MEDLINE database and not from other sources such as Scopus, Web of Science, Google Scholar, PsycINFO (psychology), CINAHL (nursing and allied health), or HMIC (healthcare management, administration, and policy). MEDLINE was selected because it is the only freely accessible primary source for interrogating the biomedical literature without requiring an institutional user account. While MEDLINE focuses primarily on medicine and the biomedical sciences, other databases cover broader disciplines. Inclusion of these databases would have been ideal, but resource constraints prevented manual review for relevance. Second, a comparison of patient outcomes between university and non-university hospitals was not performed as we were unable to determine whether the principal investigator at teaching hospitals was HA-employed or university-affiliated. Third, only crude mortality rates and LOS were evaluated. A more comprehensive review of public healthcare system key performance indicators—such as risk-adjusted or standardised mortality rates, symptom-to-intervention durations, incremental cost-effectiveness ratios, and patient satisfaction survey results—would have provided greater insight if such data were available.6 43 Important confounding factors were also not assessed, including each hospital’s annual operational income; differences in catchment population size and demographics; and variations in the scope of acute clinical services provided. For example, some institutions are recognised as level-one trauma centres or infectious disease centres. Finally, clinical research from specialties such as psychiatry and family medicine was likely under-represented, as most clinicians in these fields work in dedicated psychiatric hospitals or general outpatient clinics, which are outside the scope of this study focused on general acute care hospitals.

This study revealed that clinicians in Hong Kong’s public healthcare system produced nearly one-third of the original peer-reviewed clinical research articles published from the territory. Although the majority of these articles were case reports or retrospective studies, they achieved a relatively high degree of research influence within their respective medical specialties. Research productivity appears to be associated with improved patient outcomes, particularly in terms of crude mortality rates and 30-day unplanned readmissions. Future studies using more refined key performance indicator endpoints and adjustments for confounding factors are necessary to ascertain whether research-active institutions consistently deliver better patient outcomes.

Concept or design: PYM Woo, DKK Wong, YF Lau.
Acquisition of data: PYM Woo, DKK Wong, QHW Wong, CKL Leung, YHK Ip, DM Au, TSF Shek, B Siu, C Cheung, K Shing, A Wong, Y Khare, OWK Tsui, NHY Tang, KM Kwok, MK Chiu.
Analysis or interpretation of data: PYM Woo, DKK Wong.
Drafting of the manuscript: PYM Woo, DKK Wong.
Critical revision of the manuscript for important intellectual content: PYM Woo, DKK Wong, KHM Wan, WC Leung.

All authors had full access to the data, contributed to the study, approved the final version for publication, and take responsibility for its accuracy and integrity.

All authors have disclosed no conflicts of interest.

The authors thank Kwong Wah Hospital’s Clinical Research Centre, The Hong Kong Student Association of Neuroscience and the Hong Kong Olympia Academy for Clinical Neuroscience Research for providing essential secretarial support and assisting with data acquisition.

This research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors.

The research was approved by the Joint Chinese University of Hong Kong–New Territories East Cluster Clinical Research Ethics Committee, Hong Kong (Ref No.: 2025.256).

1. Djulbegovic B, Guyatt GH. Progress in evidence-based medicine: a quarter century on. Lancet 2017;390:415-23. Crossref

2. Ioannidis JP. Why most clinical research is not useful. PLoS Med 2016;13:e1002049. Crossref

3. Chin WY, Wong WC, Yu EY. A survey exploration of the research interests and needs of family doctors in Hong Kong. Hong Kong Pract 2019;41:29-38.

4. Ozdemir BA, Karthikesalingam A, Sinha S, et al. Research activity and the association with mortality. PLoS One 2015;10:e0118253. Crossref

5. Bennett WO, Bird JH, Burrows SA, Counter PR, Reddy VM. Does academic output correlate with better mortality rates in NHS trusts in England? Public Health 2012;126 Suppl 1:S40-3. Crossref

6. García-Romero A, Escribano Á, Tribó JA. The impact of health research on length of stay in Spanish public hospitals. Res Policy 2017;46:591-604. Crossref

7. Pons J, Sais C, Illa C, et al. Is there an association between the quality of hospitals’ research and their quality of care? J Health Serv Res Policy 2010;15:204-9. Crossref

8. Clarke M, Loudon K. Effects on patients of their healthcare practitioner’s or institution’s participation in clinical trials: a systematic review. Trials 2011;12:16. Crossref

9. Boaz A, Hanney S, Jones T, Soper B. Does the engagement of clinicians and organisations in research improve healthcare performance: a three-stage review. BMJ Open 2015;5:e009415. Crossref

10. National Institute for Health and Care Research. NIHR Annual Report 2022/23. 2024. Available from: https://www.nihr.ac.uk/reports/nihr-annual-report-202223/34501. Accessed 18 Oct 2024.

11. Ciemins EL, Mollis BL, Brant JM, et al. Clinician engagement in research as a path toward the learning health system: a regional survey across the northwestern United States. Health Serv Manage Res 2020;33:33-42. Crossref

12. Cheung BM, Yau HK. Clinical therapeutics in Hong Kong. Clin Ther 2019;41:592-7. Crossref

13. Sek AC, Cheung NT, Choy KM, et al. A territory-wide electronic health record—from concept to practicality: the Hong Kong experience. Stud Health Technol Inform 2007;129:293-6.

14. Kong X, Yang Y, Gao J, et al. Overview of the health care system in Hong Kong and its referential significance to mainland China. J Chin Med Assoc 2015;78:569-73. Crossref

15. Our Hong Kong Foundation; Hong Kong Science and Technology Parks. Developing Hong Kong into Asia’s leading clinical innovation hub. November 2023. Available from: https://ourhkfoundation.org.hk/sites/default/files/media/pdf/OHKF_BioTech_Report_2023_EN.pdf. Accessed 25 Oct 2024.

16. Kovalchik S. RISmed: download content from NCBI Databases. R package version 2.3.0 2021. Available from: https://CRAN.R-project.org/package=RISmed. Accessed 2 Jan 2023.

17. The R Foundation. R: a language and environment for statistical computing Vienna, Austria; 2022. Available from: https://www.R-project.org/. Accessed 2 Jan 2023.

18. Aquilina J, Neves JB, Tran MG. An overview of study designs. Br J Hosp Med (Lond) 2020;81:1-6. Crossref

19. Page MJ, McKenzie JE, Bossuyt PM, et al. The PRISMA 2020 statement: an updated guideline for reporting systematic reviews. Syst Rev 2021;10:89. Crossref

20. Krnic Martinic M, Meerpohl JJ, von Elm E, Herrle F, Marusic A, Puljak L. Attitudes of editors of core clinical journals about whether systematic reviews are original research: a mixed-methods study. BMJ Open 2019;9:e029704. Crossref

21. Garfield E. The history and meaning of the journal impact factor. JAMA 2006;295:90-3. Crossref

22. Suelzer EM, Jackson JL. Measures of impact for journals, articles, and authors. J Gen Intern Med 2022;37:1593-7. Crossref

23. Bergstrom CT, West JD. Assessing citations with the Eigenfactor metrics. Neurology 2008;71:1850-1. Crossref

24. Hutchins BI, Yuan X, Anderson JM, Santangelo GM. Relative Citation Ratio (RCR): a new metric that uses citation rates to measure influence at the article level. PLoS Biol 2016;14:e1002541. Crossref

25. Casadevall A, Fang FC. Impacted science: impact is not importance. mBio 2015;6:e01593-15. Crossref

26. Kreiner G. The slavery of the h-index—measuring the unmeasurable. Front Hum Neurosci 2016;10:556. Crossref

27. Reddy V, Gupta A, White MD, et al. Assessment of the NIH-supported relative citation ratio as a measure of research productivity among 1687 academic neurological surgeons. J Neurosurg 2020;134:638-45. Crossref

28. Didzbalis CJ, Avery Cohen D, Herzog I, Park J, Weisberger J, Lee ES. The Relative Citation Ratio: a modern approach to assessing academic productivity within plastic surgery. Plast Reconstr Surg Glob Open 2022;10:e4564. Crossref

29. Gupta A, Meeter A, Norin J, Ippolito JA, Beebe KS. The Relative Citation Ratio (RCR) as a novel bibliometric among 2511 academic orthopedic surgeons. J Orthop Res 2023;41:1600-6. Crossref

30. Balinska MA, Watts RA. The value of case reports in democratising evidence from resource-limited settings: results of an exploratory survey. Health Res Policy Syst 2020;18:84. Crossref

31. Suvvari TK. Are case reports valuable? Exploring their role in evidence-based medicine and patient care. World J Clin Cases 2024;12:5452-5. Crossref

32. Selby P, Autier P. The impact of the process of clinical research on health service outcomes. Ann Oncol 2011;22 Suppl 7:vii5-9. Crossref

33. Majumdar SR, Roe MT, Peterson ED, Chen AY, Gibler WB, Armstrong PW. Better outcomes for patients treated at hospitals that participate in clinical trials. Arch Intern Med 2008;168:657-62. Crossref

34. Rich AL, Tata LJ, Free CM, et al. How do patient and hospital features influence outcomes in small-cell lung cancer in England? Br J Cancer 2011;105:746-52. Crossref

35. Rochon J, du Bois A. Clinical research in epithelial ovarian cancer and patients’ outcome. Ann Oncol 2011;22 Suppl 7:vii16-9. Crossref

36. Janni W, Kiechle M, Sommer H, et al. Study participation improves treatment strategies and individual patient care in participating centers. Anticancer Res 2006;26:3661-7.

37. Downing A, Morris EJ, Corrigan N, et al. High hospital research participation and improved colorectal cancer survival outcomes: a population-based study. Gut 2017;66:89-96. Crossref

38. DiDiodato G, DiDiodato JA, McKee AS. The research activities of Ontario’s large community acute care hospitals: a scoping review. BMC Health Serv Res 2017;17:566. Crossref

39. Senecal JB, Metcalfe K, Wilson K, Woldie I, Porter LA. Barriers to translational research in Windsor Ontario: a survey of clinical care providers and health researchers. J Transl Med 2021;19:479. Crossref

40. Gehrke P, Binnie A, Chan SP, et al. Fostering community hospital research. CMAJ 2019;191:E962-6. Crossref

41. Kiparoglou V, Brown LA, McShane H, Channon KM, Shah SG. A large National Institute for Health Research (NIHR) Biomedical Research Centre facilitates impactful cross-disciplinary and collaborative translational research publications and research collaboration networks: a bibliometric evaluation study. J Transl Med 2021;19:483. Crossref

42. Burkinshaw P, Bryant LD, Magee C, et al. Ten years of NIHR research training: perceptions of the programmes: a qualitative interview study. BMJ Open 2022;12:e046410. Crossref

43. Jonker L, Fisher SJ, Dagnan D. Patients admitted to more research-active hospitals have more confidence in staff and are better informed about their condition and medication: results from a retrospective cross-sectional study. J Eval Clin Pract 2020;26:203-8. Crossref

The severe acute respiratory syndrome coronavirus 2 pandemic has been a global health crisis, resulting in substantial morbidity and mortality worldwide, with over 13 billion vaccine doses administered.1 To mitigate the risk of breakthrough infections by dominant Omicron variants, a third-dose booster following two doses of BNT162b2 vaccine (BioNTech-Pfizer, Mainz, Germany) has been rolled out. Compared with a two-dose schedule, a third dose significantly reduces the risk of infection, hospitalisation, and severe disease.2 3 However, waning anti-Omicron neutralising antibody and T cell responses have been reported even after the booster dose,4 and sustained long-term immunogenicity remains uncertain.

Advanced machine learning (ML) algorithms, such as random forest, artificial neural network (NN), and gradient boosting, have been increasingly utilised to develop prognostic models that can identify individuals at high risk of coronavirus disease 2019 (COVID-19). These models offer potential to improve risk stratification and inform targeted prevention and intervention strategies. Numerous studies have demonstrated the development of such models, which integrate various clinical, demographic, and routine laboratory variables to predict risks of COVID-19, hospitalisation, and mortality.5 6 7 8 9 However, these previous studies did not stratify patients by vaccination status, leading to heterogeneous cohorts of both vaccinated and unvaccinated individuals. This may introduce limitations and biases in model performance, given that vaccination status can substantially affect COVID-19 risk and disease severity.10 11

This study focused on individuals who had received three doses of BNT162b2, aiming to identify the ML algorithm with optimal performance for predicting COVID-19 risk using clinically available data. We also sought to identify key predictors used by the model to stratify individuals who may be more susceptible to COVID-19 despite vaccination.

This multi-centre, prospective cohort study recruited individuals aged 18 years or above who had received three doses of BNT162b2 vaccine from three vaccination centres in Hong Kong, namely, Sun Yat Sen Memorial Park Sports Centre, Queen Mary Hospital, and Sai Ying Pun Jockey Club Polyclinic, between May and August 2021. Participants volunteered for the study after being informed through flyers and announcements at the vaccination sites. All participants were screened by a trained research assistant using a checklist form (online Appendix) to confirm no active COVID-19 case or a history of the disease. Exclusion criteria included prior COVID-19 infection identified through serological testing for antibodies to the nucleocapsid protein of severe acute respiratory syndrome coronavirus 2, gastrointestinal surgery, inflammatory bowel disease, immunocompromised status (including post-transplantation, use of immunosuppressants, or receipt of chemotherapy), other medical conditions (malignancy, haematological, rheumatological or autoimmune diseases), and fewer than 14 days between the booster dose and either the study endpoint or the date of COVID-19 diagnosis.

Demographic and clinical information—including age, sex, body mass index (BMI), waist-to-hip ratio, smoking status, alcohol use, co-morbidities (hypertension, diabetes mellitus, and prediabetes), and recent medication use within 6 months of vaccination (proton pump inhibitors, statins, metformin, antibiotics,12 antidepressants, steroids, probiotics or prebiotics)—was collected. Additional data included blood pressure; blood test results (complete blood count, liver and renal function tests,13 glycated haemoglobin [HbA1c] level, lipid profile, and presence of hepatitis B surface antigen); controlled attenuation parameter (CAP) to measure liver fat14; and liver stiffness measured by transient elastography15 using FibroScan (Echosens, Paris, France). We also cross-checked the Hospital Authority’s database (eg, Clinical Management System) to verify participants’ co-morbidity conditions.

The primary outcome was COVID-19. All participants were prospectively followed from the date of their third vaccine dose until either a COVID-19 diagnosis or the end of the study (18 May 2022), whichever occurred first. Monthly follow-ups were conducted via phone calls or messages to inquire about participants’ COVID-19 status, especially during the fifth COVID-19 outbreak in Hong Kong in early 2022,16 when face-to-face meetings were not recommended. Participants were also instructed to notify the study team if they tested positive. COVID-19 diagnosis was based on self-reported symptoms followed by either a rapid antigen test or deep throat saliva reverse transcription polymerase chain reaction test.

This was a binary classification task using supervised learning algorithms, aiming to predict COVID-19 status after three vaccine doses. Predicted outcomes were labelled as ‘0’ (negative) or ‘1’ (positive). The dataset was randomly divided into training and validation sets in a 6:4 ratio.

Data preprocessing included three steps: missing data imputation, feature engineering, and data transformation. First, variables with more than 20% missing data were dropped because high levels of missingness can hinder the accuracy and reliability of imputation methods.17 18 Remaining missing values were imputed using the MICE (Multivariate Imputation by Chained Equations) package in R software (version 4.2.1, R Foundation for Statistical Computing, Vienna, Austria). Second, new features were extracted from existing variables (ie, transforming numerical variables into categorical groups and combining similar variables). Third, continuous variables were standardised through centring and scaling, whereas categorical variables were processed using one-hot encoding to ensure data compatibility for different ML algorithms.

Feature selection involved correlation analysis between variables and the dependent variable, the Boruta package in R,19 literature review, and expert consultation. A total of 37 variables were selected and ranked based on their overall importance using the aforementioned methods. Male sex, age ≥60 years, hepatitis B virus surface antigen positivity, diabetes/prediabetes, and recent medication use (antibiotics, proton pump inhibitors, probiotics/prebiotics, metformin, statins) were regarded as categorical variables (online supplementary Table 1).

Six frequently used supervised ML models were selected: logistic regression, linear discriminant analysis, random forest, naïve Bayes, NN, and extreme gradient boosting (XGBoost) [online supplementary Table 2]. Due to the imbalance in the dataset, with relatively few COVID-19 cases, multiple models were explored to assess different strategies for handling class imbalance. Hyperparameter tuning was performed using the caret package in R with grid search (3^p grid size, where p represents the number of hyperparameters) and three-fold cross-validation. The dataset was divided into three equal subsets: the model trained on two subsets and validated on the third; the process was repeated five times, with the validation subset rotated each time. Hyperparameters yielding the highest area under the receiver operating characteristic curve (AUC) on the validation set were selected. A loop function was implemented to iteratively train the model while removing a single variable from the end of the ranked list of variables. By evaluating model performance with different variable combinations, we identified the most predictive variables.

A sensitivity analysis was conducted by excluding variables not routinely available in clinical practice (eg, CAP and liver stiffness).

To compare the performance of the ML models, we calculated AUCs and used DeLong’s test to assess statistical significance among the AUCs. We estimated the best cut-off point for each model using the Youden index, selecting the threshold that maximised the sum of sensitivity and specificity. Using these cut-off points, we calculated performance metrics including sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), positive likelihood ratio (PLR), and negative likelihood ratio (NLR) to identify the best model. We also compared the miss rate (false negative rate) across models. Given the imbalanced nature of the dataset, precision-recall curves and F1 scores were used. Higher F1 scores indicate better balance between precision and recall.

All statistical analyses were conducted using R, with packages such as caret, randomForest, naivebayes, nnet, xgboost, pROC, and SHAPforxgboost for model building, evaluation, and interpretation. The DTComPair package was used to compare performance metrics.20 Continuous variables were summarised as medians and interquartile ranges (IQRs), with comparisons performed using the Wilcoxon rank-sum test. Categorical variables were presented as counts and percentages, and compared using Pearson’s Chi squared test or Fisher’s exact test, applying Bonferroni correction for multiple comparisons. SHapley Additive exPlanations (SHAP) analysis was utilised to interpret complex models by generating SHAP values to determine feature impact.

A total of 304 three-dose BNT162b2 recipients were identified between May and August 2021 (Fig 1a). The median age was 50.9 years (IQR=43.6-57.8), and 94 participants (30.9%) were men. Over a median follow-up of 2.6 months (IQR=1.8-3.1; up to 5.1 months), 27 participants (8.9%) tested positive for COVID-19. The dataset was randomly split into training and testing sets, comprising 184 (60.5%) and 120 (39.5%) participants, respectively. Table 1 summarises baseline characteristics, stratified by the outcome of interest (COVID-19 status) and by training and testing sets. Baseline characteristics prior to imputation are shown in online supplementary Table 3.

The COVID-19-positive patients had worse medical conditions than those tested negative. Specifically, they were older (with a higher proportion aged ≥60 years: 22.2% vs 16.2%), predominantly male (33.3% vs 30.7%), had greater liver fat content (median CAP: 249.0 dB/m vs 227.0 dB/m), and were more frequently diagnosed with prediabetes/diabetes (55.6% vs 38.3%). Both the training and testing sets had a comparable proportion of COVID-19–positive cases (8.3-9.2%). Most independent variables were similarly distributed between sets (P>0.05), although CAP differed significantly (Table 1).

We trained six different ML algorithms on the training set to predict COVID-19. Model performance was evaluated using three-fold cross-validation (Fig 1b). Concerning the testing set, performance metrics for each model are reported in Table 2 and AUCs are summarised in Figure 2. A comparison of AUCs between training and testing sets for the six ML models is presented in online supplementary Figure 1. All models showed a slight decrease in AUC from the training to the testing set, indicating some degree of overfitting. Notably, the NN model did not exhibit a significant AUC reduction, suggesting it was less susceptible to overfitting than other models.

Of the six ML models evaluated, the NN algorithm performed best (AUC: 0.74, 95% CI=0.60-0.88), followed by XGBoost (AUC: 0.62, 95% CI=0.42-0.82) [Fig 2]. Using the optimal cut-off value estimated by the maximum Youden index, performance metrics are summarised in Table 2. The 2×2 confusion matrix tables, which summarise the numbers of true positives, true negatives, false positives, and false negatives for each model’s predictions, are shown in online supplementary Table 4. Multiple comparisons between the NN and other models in terms of performance metrics are presented in online supplementary Table 5.

The NN and linear discriminant analysis models achieved the highest sensitivity, with values of 90% (95% CI=55%-100%) and 80% (95% CI=44%-97%), respectively. The random forest model had the best specificity (72%, 95% CI=62%-80%). The NN model also had the highest NPV (98%, (95% CI=92%-100%) and the best likelihood ratios (PLR: 2.15, 95% CI=1.59-2.91; NLR: 0.17, 95% CI=0.03-1.11) [Table 2]. It classified 45.8% of participants as high risk for COVID-19, with a miss rate or false negative rate of 10% (Table 3). Precision-recall curves and F1 scores for all models are shown in online supplementary Figure 2, offering a more precise evaluation of model performance in the context of an imbalanced dataset. With a precision baseline of 0.092, naïve Bayes and random forest models recorded AUC values of around 0.10, reflecting modest discrimination ability under class imbalance. The NN model achieved an F1 score of 0.277, highlighting a better balance between precision and recall.

According to the best-performing model (the NN model), the five most important predictors of COVID-19 risk were CAP, alanine aminotransferase level, age (≥60 years), presence of prediabetes/diabetes, and aspartate aminotransferase (AST) level, with relative importance values of 14.9%, 10.1%, 9.4%, 8.4%, and 7.9%, respectively (Fig 3). These were further confirmed by SHAP analysis, a method specifically compatible with ensemble algorithms (ie, XGBoost) that quantifies the contribution of each input variable to the model’s prediction. When SHAP analysis was applied to the second best-performing model (XGBoost), leading variables remained similar to those in the NN model, except for BMI which ranked highest in importance (with a mean absolute SHAP value of 0.992) in the XGBoost model (online supplementary Fig 3). The SHAP analysis in online supplementary Figure 3b also provided deeper insights into the contribution of each variable to the model’s prediction. Among leading variables in the XGBoost model, higher CAP (red dots), lower BMI (blue dots), and age ≥60 years (red dots) had a positive impact (right side of the plot) on COVID-19 prediction. In terms of high-density lipoprotein (HDL) and AST levels, the SHAP plot showed a wide distribution with mixed colours, suggesting that HDL and AST levels had diverse impacts on COVID-19 prediction.

Excluding CAP and liver stiffness, XGBoost achieved the best performance (AUC: 0.66, 95% CI=0.50-0.82), followed by naïve Bayes, logistic regression, linear discriminant analysis, random forest, and NN models (AUCs: 0.49- 0.63) [online supplementary Fig 4]. The top five predictors in the XGBoost model were BMI, alanine aminotransferase level, HDL level, HbA1c level, and age ≥60 years (online supplementary Fig 5). In the NN model, the top predictors were AST level, HbA1c level, HDL level, hepatitis B virus antigen positivity, and alanine aminotransferase level (online supplementary Fig 6).

In this study involving three-dose BNT162b2 recipients, the NN model achieved satisfactory performance in predicting COVID-19 using baseline clinical data. The leading predictors identified were age ≥60 years, presence of prediabetes/diabetes, CAP, alanine aminotransferase level, and AST level, highlighting the need for vigilance among fully vaccinated individuals, especially those with concomitant co-morbidities.

Advanced age, prediabetes/diabetes, and abnormal liver condition (ie, high fatty liver content and abnormal liver function test results) were significant predictors of high infection risk, consistent with previous studies.21 22 23 24 25 A meta-analysis of 18 studies revealed a higher prevalence of diabetes (11.5%) among hospitalised COVID-19 patients21 compared to the general population (9.3%).26 Studies have found that the presence of preexisting diabetes or hyperglycaemia is associated with higher risks of severe illness, mortality, and complications in COVID-19 patients.22 23 This elevated risk is likely due to impaired immune function, chronic inflammation, and common cardiovascular and metabolic co-morbidities in diabetic patients.27 28 Individuals with liver diseases or abnormal liver function test results also exhibit higher risks of severe COVID-19 and complications.24 25

This study is among the few that have developed ML models to predict COVID-19 in recipients of three doses of BNT162b2. No prior studies have developed COVID-19 prognostic models with clear information on vaccination status, type, and number of doses. A study from Hong Kong11 showed that a timely third vaccine dose strongly protected against Omicron BA.2 variant infections, the dominant strain in Hong Kong during our study period. The effectiveness of vaccination against infection declined over time after two doses but was restored to a high level after a third dose, resulting in significantly lower risks of infection, hospitalisation, and severe illness compared with those who received only two doses.2 3 By including only three-dose vaccinated patients in the development of ML models, the resulting models may be more accurate in predicting COVID-19 risk and severity among vaccinated individuals. This can be particularly important in settings where vaccination rates are high and breakthrough infections are a concern; it may help identify individuals with higher infection risk who could benefit from additional precautions or interventions.

Our study offers practical value by enabling risk stratification, allowing healthcare providers to focus resources on higher-risk populations. It informs public health strategies by identifying factors associated with increased COVID-19 risk despite vaccination, guiding targeted campaigns and resource allocation. Additionally, an understanding of risk predictors in vaccinated individuals supports tailored booster strategies. The identification of key variables such as age, prediabetes/diabetes, and liver enzyme levels also encourages further research into underlying mechanisms and potential interventions.

However, this study had some limitations. First, the small sample size (~300 participants) may affect model performance and generalisability. The dataset size was constrained by specific inclusion criteria, but this represented the maximum size available for model training. We believe that selection of high-quality data maximises training efficacy. Second, we did not include gut microbiota data, which may be associated with COVID-19 vaccine immunogenicity.29 A focus on readily available clinical data facilitates practical and clinically relevant predictive models. Third, our dataset exhibited significant class imbalance, such that only 8.9% of participants developed COVID-19 within 6 months. Whereas receiver operating characteristic curve analysis provides an optimistic assessment, we also used precision-recall curves and F1 scores for a more realistic evaluation. Fourth, although missing values for certain variables might introduce error into the prediction models, the small percentage of missing data and the use of multiple imputation likely had minimal impact on model accuracy. Fifth, COVID-19 cases were self-reported and confirmed by either rapid antigen or polymerase chain reaction tests. In Hong Kong, rapid antigen tests have a false negative rate of approximately 15% (sensitivity: 85%)30 but a high specificity of 99.93%,30 indicating very few false positives. Although some cases may have gone unreported or untested, we believe that the majority adhered to testing requirements as mandated by law. Additionally, we did not grade infection severity, and there were no hospitalised cases in our cohort, limiting our ability to predict hospitalisation outcomes in this study. Sixth, the NN model—our best-performing model—is complex and has low interpretability. We used a variable importance plot to visualise and identify the most influential features, enhancing its practical application. It should be noted that the other models demonstrated suboptimal performance, with AUCs below 0.7. The NN model’s superior performance is likely due to its ability to capture complex patterns and interactions. Simpler models struggled with the dataset’s complexity, class imbalance, non-linear relationships, and outliers. Finally, although this study offers insights into the use of advanced ML models to predict COVID-19 outcomes, its generalisability is limited. Overfitting remains a concern despite mitigation techniques (eg, regularisation, pruning, and ensemble methods). The complexity of our models and the dataset hinder generalisability. Variability in vaccines, booster intervals, doses, demographics, and study design further impacts the generalisability of our model. Future studies should include diverse populations and vaccine types to enhance applicability. External validation of our results in other centres is also warranted.

The NN model is a useful tool for identifying individuals at low risk of COVID-19 within 6 months after receiving three doses of BNT162b2. Key features selected by the model highlight the central role of metabolic risk factors (prediabetes/diabetes, non-alcoholic fatty liver disease, and steatohepatitis) in vaccine immunogenicity.

Concept or design: JT Tan, KS Cheung.
Acquisition of data: JT Tan, R Zhang, KH Chan.
Analysis or interpretation of data: JT Tan, KS Cheung.
Drafting of the manuscript: JT Tan.
Critical revision of the manuscript for important intellectual content: KS Cheung, IFN Hung.

All authors had full access to the data, contributed to the study, approved the final version for publication, and take responsibility for its accuracy and integrity.

All authors have disclosed no conflicts of interest.

This research was funded by the Health and Medical Research Fund of the former Food and Health Bureau, Hong Kong SAR Government (Ref No.: COVID1903010, Project 16). The funder had no role in the study design, data collection/analysis/interpretation, or manuscript preparation.

The research was approved by the Institutional Review Board of The University of Hong Kong/Hospital Authority Hong Kong West Cluster, Hong Kong (Ref No.: UW 21-216). Participants provided written informed consent to participate in this study.

The supplementary material was provided by the authors and some information may not have been peer reviewed. Accepted supplementary material will be published as submitted by the authors, without any editing or formatting. Any opinions or recommendations discussed are solely those of the author(s) and are not endorsed by the Hong Kong Academy of Medicine and the Hong Kong Medical Association. The Hong Kong Academy of Medicine and the Hong Kong Medical Association disclaim all liability and responsibility arising from any reliance placed on the content.

1. World Health Organization. WHO Coronavirus (COVID-19) Dashboard. 2023. Available from: https://data.who.int/dashboards/covid19/cases. Accessed 4 May 2023.

2. Andrews N, Stowe J, Kirsebom F, et al. Effectiveness of COVID-19 booster vaccines against COVID-19–related symptoms, hospitalization and death in England. Nat Med 2022;28:831-37. Crossref

3. Andrews N, Stowe J, Kirsebom F, et al. COVID-19 vaccine effectiveness against the Omicron (B.1.1.529) variant. N Engl J Med 2022;386:1532-46. Crossref

4. Peng Q, Zhou R, Wang Y, et al. Waning immune responses against SARS-CoV-2 variants of concern among vaccinees in Hong Kong. EBioMedicine 2022;77:103904. Crossref

5. Willette AA, Willette SA, Wang Q, et al. Using machine learning to predict COVID-19 infection and severity risk among 4510 aged adults: a UK Biobank cohort study. Sci Rep 2022;12:7736. Crossref

6. Wynants L, Van Calster B, Collins GS, et al. Prediction models for diagnosis and prognosis of COVID-19: systematic review and critical appraisal. BMJ 2020;369:m1328. Crossref

7. Subudhi S, Verma A, Patel AB, et al. Comparing machine learning algorithms for predicting ICU admission and mortality in COVID-19. NPJ Digit Med 2021;4:87. Crossref

8. Brinati D, Campagner A, Ferrari D, Locatelli M, Banfi G, Cabitza F. Detection of COVID-19 infection from routine blood exams with machine learning: a feasibility study. J Med Syst 2020;44:135. Crossref

9. Yao H, Zhang N, Zhang R, et al. Severity detection for the coronavirus disease 2019 (COVID-19) patients using a machine learning model based on the blood and urine tests. Front Cell Dev Biol 2020;8:683. Crossref

10. Baden LR, El Sahly HM, Essink B, et al. Efficacy and safety of the mRNA-1273 SARS-CoV-2 vaccine. N Engl J Med 2021;384:403-16. Crossref

11. Zhou R, Liu N, Li X, et al. Three-dose vaccination–induced immune responses protect against SARS-CoV-2 Omicron BA.2: a population-based study in Hong Kong. Lancet Reg Health West Pac 2023;32:100660. Crossref

12. Cheung KS, Lam LK, Zhang R, et al. Association between recent usage of antibiotics and immunogenicity within six months after COVID-19 vaccination. Vaccines (Basel) 2022;10:1122. Crossref

13. Cheung KS, Mok CH, Mao X, et al. COVID-19 vaccine immunogenicity among chronic liver disease patients and liver transplant recipients: a meta-analysis. Clin Mol Hepatol 2022;28:890-911. Crossref

14. Cheung KS, Lam LK, Hui RW, et al. Effect of moderate-to-severe hepatic steatosis on neutralising antibody response among BNT162b2 and CoronaVac recipients. Clin Mol Hepatol 2022;28:553-64. Crossref

15. Cheung KS, Lam LK, Mao X, et al. Effect of moderate to severe hepatic steatosis on vaccine immunogenicity against wild-type and mutant virus and COVID-19 infection among BNT162b2 recipients. Vaccines (Basel) 2023;11:497. Crossref

16. Cheung PH, Chan CP, Jin DY. Lessons learned from the fifth wave of COVID-19 in Hong Kong in early 2022. Emerg Microbes Infect 2022;11:1072-8. Crossref

17. Little RJ, Rubin DB. Statistical Analysis with Missing Data, 3rd edition. New York [NY]: John Wiley & Sons; 2019. Crossref

18. Dong Y, Peng CY. Principled missing data methods for researchers. Springerplus 2013;2:222. Crossref

19. Kursa MB, Rudnicki WR. Feature selection with the Boruta package. J Stat Softw 2010;36:1-13. Crossref

20. Kitcharanant N, Chotiyarnwong P, Tanphiriyakun T, et al. Development and internal validation of a machine-learning–developed model for predicting 1-year mortality after fragility hip fracture. BMC Geriatr 2022;22:451. Crossref

21. Singh AK, Gillies CL, Singh R, et al. Prevalence of co-morbidities and their association with mortality in patients with COVID-19: a systematic review and meta-analysis. Diabetes Obes Metab 2020;22:1915-24. Crossref

22. Zhu L, She ZG, Cheng X, et al. Association of blood glucose control and outcomes in patients with COVID-19 and pre-existing type 2 diabetes. Cell Metab 2020;31:1068-77.e3. Crossref

23. Yang JK, Feng Y, Yuan MY, et al. Plasma glucose levels and diabetes are independent predictors for mortality and morbidity in patients with SARS. Diabet Med 2006;23:623-8. Crossref

24. Singh S, Khan A. Clinical characteristics and outcomes of coronavirus disease 2019 among patients with preexisting liver disease in the United States: a multicenter research network study. Gastroenterology 2020;159:768-771.e3. Crossref

25. Simon TG, Hagström H, Sharma R, et al. Risk of severe COVID-19 and mortality in patients with established chronic liver disease: a nationwide matched cohort study. BMC Gastroenterol 2021;21:439. Crossref

26. Saeedi P, Petersohn I, Salpea P, et al. Global and regional diabetes prevalence estimates for 2019 and projections for 2030 and 2045: results from the International Diabetes Federation Diabetes Atlas, 9th edition. Diabetes Res Clin Pract 2019;157:107843. Crossref

27. Pal R, Bhadada SK. COVID-19 and diabetes mellitus: an unholy interaction of two pandemics. Diabetes Metab Syndr 2020;14:513-7. Crossref

28. Azar WS, Njeim R, Fares AH, et al. COVID-19 and diabetes mellitus: how one pandemic worsens the other. Rev Endocr Metab Disord 2020;21:451-63. Crossref

29. Ng HY, Leung WK, Cheung KS. Association between gut microbiota and SARS-CoV-2 infection and vaccine immunogenicity. Microorganisms 2023;11:452. Crossref

30. Zee JS, Chan CT, Leung AC, et al. Rapid antigen test during a COVID-19 outbreak in a private hospital in Hong Kong. Hong Kong Med J 2022;28:300-5. Crossref

Lung cancer is the leading cause of cancer-related deaths in Hong Kong, with a high incidence and low 5-year survival rate.1 Early diagnosis and treatment are crucial to improving survival outcomes, as demonstrated in lung cancer screening trials and real-world experiences.2 3 4 5 To date, low-dose computed tomography (CT) screening has not been implemented in Hong Kong.6 However, the widespread clinical use of CT has led to increased detection of lung nodules on thoracic CT scans.7 Despite general awareness of lung nodule management guidelines, adherence is often suboptimal.8 9 Over the past few decades, an increasing number of centres have developed programmes to manage incidental lung nodules.10 This study aimed to review the clinical outcomes of a lung nodule surveillance programme at a tertiary hospital in Hong Kong. It outlines the structure of this comprehensive programme, characteristics of lung nodules, invasive interventions, stage distribution, treatment modalities, and initial survival data.

This single-centre prospective cohort study included all patients enrolled in a lung nodule surveillance programme at Queen Mary Hospital, Hong Kong, during its first year of implementation, from 1 January to 31 December 2022. The programme was launched with the following objectives: (1) to streamline the referral process and promote cohesive care among practitioners; (2) to identify high-risk lung nodules requiring timely invasive intervention; (3) to provide a multidisciplinary platform with a nurse navigator coordinating care across specialties; and (4) to establish rapport between patients and healthcare staff, thereby improving compliance. Recruitment criteria included patients under surveillance for lung nodules (lesions <3 cm) or lung masses. Patients with lung cancer receiving active treatment were excluded. Referral sources included lung cancer screening, incidental findings on CT or chest radiographs, and symptomatic presentations. All recruited cases were followed until the lung nodule was confirmed as benign or malignant, or until participants declined follow-up or died. The programme algorithm is shown in Figure 1. Upon entry into the pathway, a nurse coordinator conducts initial screening and monitors patients’ waiting time. The clinical team, comprising doctors and nurses, reviews each case to perform risk stratification, formulate investigation and follow-up plans, arrange appropriate interventions, provide patient education, and ensure timely follow-up. Patients are managed according to the 2017 Fleischner Society Guidelines11 and the Clinical Practice Consensus Guidelines for Asia.12 Lung nodules requiring intensive evaluation or tissue sampling are discussed by a multidisciplinary team that includes respiratory physicians, chest radiologists, cardiothoracic surgeons, and clinical oncologists. Nurse coordinators regularly review consultation notes and track investigation results. Overdue scans or missed follow-ups are identified and rescheduled. Surveillance continues until the lung nodule is confirmed as benign or malignant.

The demographic and clinical characteristics of all recruited individuals including sex, age, smoking history, and co-morbidities were retrieved from the Clinical Management System of the Hospital Authority. In pathology reports, cases where non–small-cell lung cancer (NSCLC) could not be further classified were recorded as NSCLC. In the results, adenocarcinoma, squamous cell carcinoma, smallcell lung cancer, and NSCLC were listed as mutually exclusive categories. Cancer staging was assigned based on pathological staging when available, or clinical staging otherwise, in accordance with the eighth edition of the International Association for the Study of Lung Cancer staging classification.13 Patients with stage I or II lung cancer were categorised as early-stage, and those with stage III or IV disease as late-stage. The STROBE (Strengthening the Reporting of Observational Studies in Epidemiology) reporting guideline was followed in the preparation of this manuscript. Primary outcomes were the proportion of lung cancer cases detected and their stage at diagnosis. Secondary outcomes included invasive investigations performed, adverse events related to invasive procedures, subgroup analysis by referral source, avoidance of overdue scans and loss to follow-up, and lung cancer treatment and survival.

Continuous variables were reported as mean with standard deviation or median with interquartile range, as appropriate. Categorical variables were reported as number and percentage. Clinical characteristics between subgroups were analysed using Student’s t test or Chi squared test. Progression-free survival (PFS) and overall survival (OS) were analysed using the Kaplan–Meier method, and factors affecting survival were evaluated using Cox regression analysis. All analyses were performed using SPSS (Windows version 23.0; IBM Corp, Armonk [NY], US).

A total of 1471 adult patients with lung nodules or lung masses were enrolled in the lung nodule surveillance programme. Demographic characteristics are summarised in online supplementary Table 1. The mean age was 68 years; 726 (49.4%) were men, 954 (64.9%) were never-smokers, 150 (10.2%) were current smokers, and 367 (24.9%) were ex-smokers. Regarding co-morbidities, 140 patients (9.5%) had chronic obstructive pulmonary disease, 35 (2.4%) had interstitial lung disease, 59 (4.0%) had asthma, and 105 (7.1%) had a history of tuberculosis. Of the 1471 participants, 291 (19.8%) underwent invasive diagnostic procedures, resulting in the diagnosis of lung cancer in 133 patients (9.0%), including 131 primary lung cancers and two cases of lung metastasis from pancreatic carcinoma (online supplementary Fig 1). Pathological findings of the 158 patients without lung cancer are presented in online supplementary Table 2. Referral sources and radiographic characteristics of 266 patients, including 133 with lung cancer and 133 matched controls (matched by age, sex, and smoking history) without lung cancer, are presented in online supplementary Table 3. Among these 266 patients, 193 (72.6%) had more than one pulmonary lesion, 67.3% of dominant nodules were solid, and 50.0% were located in the upper lobes. The median size of the dominant lung lesion was 12 mm (interquartile range, 6-26). Patients with lung cancer had significantly larger pulmonary lesions (median size: 24 mm vs 6 mm, P<0.001) and a higher proportion of part-solid nodules (21.1% vs 3.0%, P<0.001) than those without lung cancer. Referral sources among the 266 patients included lung cancer screening (29/266, 10.9%), incidental findings on CT or chest radiographs (153/266, 57.5%), and symptomatic presentation (84/266, 31.6%). There were no significant differences between groups in referral source, nodule number, or location. Among the 133 patients without lung cancer, 31 (23.3%) missed scheduled follow-up and 12 (9.0%) had overdue scans. All were identified by programme coordinators and rescheduled for follow-up.

The characteristics of patients with confirmed lung cancer (n=133), both overall and stratified by sex (men: n=69; women: n=64) and smoking history (non-smokers: n=81; current/ex-smokers: n=52), are presented in Table 1. The mean age of patients with lung cancer was 69.4 years; 51.9% were men and 60.9% were non-smokers. There were no significant differences between groups in terms of age or body mass index. Compared with male patients, female patients had significantly lower cigarette exposure (mean pack-years: 1.7 vs 25.7, P<0.001) and a higher prevalence of epidermal growth factor receptor (EGFR) mutations (66.7% vs 38.2%, P=0.008). Compared with current or ex-smokers, non-smokers were more likely to have adenocarcinoma (90.1% vs 73.1%, P=0.003), EGFR mutations (69.6% vs 22.5%, P<0.001), and stage I lung cancer (55.6% vs 32.7%, P=0.041).

Of the 133 patients with lung cancer, 92 (69.2%) had more than one pulmonary lesion. Among the dominant lesions, 60.9% were solid and 56.4% were located in the upper lobes. The median size of the dominant lesion was 24.0 mm (interquartile range, 16.0-39.3), and 48 (36.1%) cases presented with lung masses. Male patients and individuals with a history of smoking had significantly larger pulmonary lesions (mean size: 37.2 mm vs 26.1 mm for men and women, P=0.013; 38.8 mm vs 27.5 mm for current or ex-smokers and non-smokers, P=0.014). Female patients exhibited a higher prevalence of ground-glass opacities (GGOs) compared with male patients (28.1% vs 4.3%, P=0.002) and a lower proportion of solid lesions (50.0% vs 71.0%, P=0.004). Non-smokers were more likely to have pure GGOs (23.5% vs 3.8%, P=0.004) and part-solid GGOs (28.4% vs 9.6%, P=0.016) but fewer solid lesions (48.1% vs 80.8%, P<0.001), compared with current or ex-smokers.

Among the 291 patients who underwent invasive investigations, two experienced a small pneumothorax following transbronchial biopsy during bronchoscopy, and five developed a minor pneumothorax after CT-guided biopsy of a lung lesion; all resolved with conservative management. Another two patients had a large pneumothorax following CT-guided biopsy of a lung nodule, requiring chest drain insertion, which was removed after 3 days. One patient developed trace perilesional haemorrhage after CT-guided biopsy of a lung lesion, which was managed conservatively. No patients who underwent invasive procedures experienced complications such as haemothorax, respiratory failure requiring mechanical ventilation, or death. The diagnoses for the 133 patients with lung cancer were established using various diagnostic procedures: 39 (29.3%) cases via bronchoscopy with or without endobronchial ultrasound; 32 (24.1%) via CT-guided biopsy of a lung lesion; 44 (33.1%) through surgical biopsy; nine (6.8%) through cytological analysis of sputum, pleural fluid, ascites, or pericardial effusion; seven (5.3%) by ultrasound-guided fine needle biopsy of cervical lymph nodes or subcutaneous nodules; one (0.8%) via endoscopic ultrasound-guided biopsy of an adrenal lesion; and one (0.8%) through liquid biopsy detecting EGFR mutations as the patient declined tissue biopsy (online supplementary Fig 2). Of the 133 patients, 17 (12.8%) underwent invasive investigations at metastatic sites.

Among the 133 lung cancer cases, 62 (46.6%) were stage I, 10 (7.5%) were stage II, 16 (12.0%) were stage III, and 45 (33.8%) were stage IV. No cases of adenocarcinoma in situ arising from GGOs were identified among the 291 patients who underwent invasive investigations; however, five cases of minimally invasive adenocarcinoma (pT1mi, pathological stage IA1) were detected. The proportion of stage I lung cancer in the lung nodule surveillance programme (46.6%) was significantly higher than that reported by the Hong Kong Cancer Registry (HKCR) in 2021 (17.4%, P<0.001).1 Conversely, the proportion of stage IV cases was significantly lower in the surveillance programme (33.8%, 45 of 133) compared with the HKCR data (56.9%, P<0.001) [Fig 2a]. Overall, 54.1% (72 of 133) of cases in the surveillance programme were diagnosed at an early stage (stage I-II), compared with 22.4% in the HKCR data (P<0.001) [Fig 2b].

Of the 133 lung cancer patients, 60 (45.1%) underwent surgical resection; seven (5.3%) received stereotactic body radiation therapy; 27 (20.3%) were treated with systemic chemotherapy with or without radiotherapy, combined chemoimmunotherapy (chemoIO), or IO alone; and 21 (15.8%) received targeted therapy. Three patients (2.3%) received palliative radiotherapy as their first-line treatment. The remaining 12 (9.0%) patients opted for best supportive care, one (0.8%) patient pursued traditional Chinese medicine, and two moved out of Hong Kong. Two cases who left Hong Kong were lost to follow-up after diagnosis.

Compared with non-smokers, a lower proportion of patients with a history of smoking underwent surgical resection (45 of 79 [57.0%] non-smokers vs 15 of 52 [28.8%] current/ex-smokers, P=0.0014) and targeted therapy (19 of 79 [24.1%] vs 2 of 52 [3.8%], P=0.0019). In contrast, a higher proportion of current/ex-smokers received chemotherapy, chemoradiotherapy, chemoIO, or IO alone (8 of 79 [10.1%] vs 19 of 52 [36.5%], P<0.001) and best supportive care (2 of 79 [2.5%] vs 10 of 52 [19.2%], P=0.0014) [Fig 3a]. Patients diagnosed with stage I lung cancer primarily received surgical resection or curative RT, whereas those diagnosed at stage III or IV were more likely to receive systemic therapy or best supportive care (Fig 3b).

As of the survival analysis conducted in April 2024, the median survival for all patients had not been reached. Among patients with late-stage lung cancer, the median PFS was 282 days (95% confidence interval [95% CI]=52-512), and the median OS was 438 days (95% CI=137-739). Both the median disease-free survival and OS for patients with early-stage lung cancer had not been reached at the time of analysis (Fig 4).

Univariate Cox regression analysis revealed that female sex (P<0.02), age (P<0.001), smoking history (P<0.02), lung lesion size (P≤0.001), disease stage (P<0.001), and surgical resection (P<0.001) were significant predictors of both PFS and OS. Multivariate Cox regression analysis showed that male sex, age, and late-stage disease were independently associated with survival. For male sex, the hazard ratios were 2.831 (95% CI=1.066-7.517) for PFS and 2.931 (1.107-7.756) for OS (P=0.037 and 0.030, respectively). For age, the hazard ratios were 1.082 (1.040-1.136) for PFS and 1.117 (1.067-1.170) for OS, and for late-stage disease, the hazard ratios were 12.664 (4.405-36.408) and 11.791 (4.132-33.651) for PFS and OS, respectively (all P<0.001) [Table 2]. There were no statistically significant differences in survival rates based on nodule characteristics or locations.

In this single-centre study, the implementation of a lung nodule surveillance programme resulted in 54.1% of lung cancer patients being diagnosed at an early, potentially curable stage, which may contribute to improved survival outcomes. Compared with HKCR 2021 data, this reflects a 31.7% increase in the proportion of early-stage diagnoses (from 22.4% to 54.1%) and a potential 23.1% reduction in the proportion of patients diagnosed with stage IV disease (from 56.9% to 33.8%).

The prevalence of lung nodules detected in lung cancer screening trials varies considerably, with reported rates ranging from 6.8% to 50.9%.7 14 15 16 17 18 19 The National Lung Screening Trial showed that 96% of nodules detected in high-risk smokers were non-malignant.2 The Taiwan Lung Cancer Screening in Never-Smoker Trial, a multicentre cohort of 12 011 never-smokers, revealed a lung nodule prevalence of 17.4% and a cancer prevalence of 2.6% at baseline screening.20 An ongoing clinical trial in Hong Kong is investigating lung cancer screening in individuals with a smoking history or a family history of lung cancer.21 Although low-dose CT screening is a proven modality that significantly reduces lung cancer mortality in some countries, current guidelines recommend screening only for a limited high-risk population. Moreover, the detection of incidental lung nodules has increased with the widespread use of CT scans.22 In our study, such nodules accounted for 57.5% of all cases. Increased referrals for incidental nodules likely contributed to longer waiting times for follow-up appointments. Previous studies have reported delayed or failed tracking of incidental pulmonary nodules in 40% to 70% of patients,8 23 and one study reported a 27% loss to follow-up.19 A dedicated lung nodule surveillance programme can complement lung cancer screening by facilitating early detection in individuals who do not meet current screening criteria but present with incidental pulmonary nodules. In our study, overdue scans (9.0%) and missed follow-up appointments (23.3%) were promptly identified and rescheduled by programme coordinators, ensuring close monitoring and timely intervention to optimise patient outcomes.

Lung nodule surveillance programmes have shown success across various settings. For example, the lung nodule registry at National Jewish Health established a database of patients with identified nodules, resulting in improved follow-up imaging and a stage shift towards earlier lung cancer diagnosis.24 25 Similarly, a comprehensive programme in Tennessee in the US reported an increase in the proportion of stage I or II cancer diagnoses from 23% to 38% following the implementation of electronic and manual chart reviews.10 The lung nodule surveillance programme in this study not only actively tracked reports but also provided clinical assessments, estimated patient cancer risk, and optimised guideline adherence. Our findings echo those of previous studies: stage I and II lung cancer was detected in 54.1% of cases, compared with 22.4% reported in the 2021 HKCR report.

Smoking remains a significant risk factor for lung cancer in both men and women.26 Our study showed that smokers presented with larger nodules, more advanced disease stages, and a lower likelihood of undergoing surgical resection compared with non-smokers. Unlike Western populations where 10% to 20% of lung cancer patients are non-smokers,27 approximately 38% of lung cancer cases in Asia occur in non-smokers.2 18 28 29 30 31 32 33 34 35 Our findings reflect this regional trend, with 60.9% of confirmed lung cancer patients being non-smokers; among them, 38.3% were diagnosed at a late stage. Non-smokers in our cohort were more likely to have GGOs, adenocarcinoma, EGFR mutations, and early-stage lung cancer at diagnosis—findings consistent with previous studies.23 35 36 37 These results suggest that lung cancer control efforts in Hong Kong should target both smokers and non-smokers with relevant risk factors.5 Beyond smoking status, emerging research highlighted sex-based differences in lung cancer biology, including hormonal factors, molecular changes, genetic predispositions, the presence of human papillomavirus, nontuberculous mycobacteria, and prior radiation therapy for breast cancer.35

Previous studies have shown that improved survival is associated with stage shifts.38 39 For example, Yang et al5 reported an increase in stage I disease from 15.9% to 58.8% and a decrease in stage IV disease from 60.3% to 25.7% between 2006 and 2019, during which the 5-year survival rate for lung cancer in Taiwan rose from 22.1% to 54.9%. In the present study, although the median PFS and OS had not yet been reached at the time of analysis; patients with early-stage lung cancer had significantly better survival outcomes. Multivariate analysis identified sex, stage, and age as independent predictors of survival, while no significant survival differences were observed based on nodule density or location. These findings suggest that although the physical attributes of lung nodules are important for diagnosis, survival outcomes are more strongly influenced by factors such as stage at diagnosis, patient demographics, and treatment strategies.

This study has several limitations. First, the follow-up period was relatively short at just over 15 months, and the median PFS and OS were not reached for early-stage patients. Second, the observational design of this single-centre study limits the ability to establish a causal relationship between the lung nodule surveillance programme and the observed stage shift. Although patients with clinically metastatic lung cancer were not excluded, the study cohort may not fully represent the broader lung cancer population in Hong Kong. Nevertheless, the findings provide indirect evidence that a lung nodule surveillance programme may facilitate early-stage detection. To confirm the programme’s impact on stage shift and survival outcomes, randomised controlled trials or case-control studies comparing enrolled patients with non-enrolled but eligible patients are warranted. Furthermore, local studies evaluating the cost-effectiveness of such programmes would provide further evidence to optimise care for patients with lung nodules.

This lung nodule surveillance programme improved follow-up compliance and facilitated timely, appropriate investigations for high-risk patients. Consequently, more than half of lung cancer cases were diagnosed at an early stage, enabling provision of curative treatments and potentially contributing to improved survival outcomes.

Concept or design: DCL Lam, LYW Shong.
Acquisition of data: WC Chong, LYW Shong, PI Cheang, WC Choy.
Analysis or interpretation of data: LYW Shong, FKP Chan, WC Kwok, WC Chong, MSM Ip, DCL Lam.
Drafting of the manuscript: DCL Lam and LYW Shong.
Critical revision of the manuscript for important intellectual content: DCL Lam, LYW Shong, WC Kwok.

All authors had full access to the data, contributed to the study, approved the final version for publication, and take responsibility for its accuracy and integrity.

As an editor of the journal, WC Kwok was not involved in the peer review process. Other authors have disclosed no conflicts of interest.

The authors thank all patients for their participation.

Part of the research data was presented at the Hospital Authority Convention 2024, 16-17 May 2024, Hong Kong.

The research described in this report was supported in part by the Lee and the Ho Families Respiratory Research Fund. The funder had no role in the study design, data collection/analysis/interpretation, or manuscript preparation.

This study was approved by the Institutional Review Board of The University of Hong Kong/Hospital Authority Hong Kong West Cluster, Hong Kong (Ref No.: UW23-101) and was conducted in full compliance with the ICH E6 guideline for Good Clinical Practice and the principles of the Declaration of Helsinki. The requirement for patient consent was waived by the Board as the study involved minimal risk and used anonymised data collected during routine clinical care without altering patient management.

The supplementary material was provided by the authors and some information may not have been peer reviewed. Accepted supplementary material will be published as submitted by the authors, without any editing or formatting. Any opinions or recommendations discussed are solely those of the author(s) and are not endorsed by the Hong Kong Academy of Medicine and the Hong Kong Medical Association. The Hong Kong Academy of Medicine and the Hong Kong Medical Association disclaim all liability and responsibility arising from any reliance placed on the content.

1. Hospital Authority. Hong Kong Cancer Registry 2021. Available from: https://www3.ha.org.hk/cancereg/. Accessed 19 Apr 2024.

2. Aberle DR, Adams AM, Berg CD, et al. Reduced lung-cancer mortality with low-dose computed tomographic screening. N Engl J Med 2011;365:395-409. Crossref

3. de Koning HJ, van der Aalst CM, de Jong PA, et al. Reduced lung-cancer mortality with volume CT screening in a randomized trial. N Engl J Med 2020;382:503-13. Crossref

4. Becker N, Motsch E, Trotter A, et al. Lung cancer mortality reduction by LDCT screening—results from the randomized German LUSI trial. Int J Cancer 2020;146:1503-13. Crossref

5. Yang CY, Lin YT, Lin LJ, et al. Stage shift improves lung cancer survival: real-world evidence. J Thorac Oncol 2023;18:47-56. Crossref

6. Centre for Health Protection, Department of Health, Hong Kong SAR Government. Cancer Expert Working Group on Cancer Prevention and Screening. Recommendations on Prevention and Screening for Lung Cancer For Health Professionals. Jun 2023. Available from: https://www.chp.gov.hk/files/pdf/lung_cancer_professional_hp.pdf. Accessed 31 Aug 2024.

7. Gould MK, Tang T, Liu IL, et al. Recent trends in the identification of incidental pulmonary nodules. Am J Respir Crit Care Med 2015;192:1208-14. Crossref

8. Shelver J, Wendt CH, McClure M, et al. Effect of an automated tracking registry on the rate of tracking failure in incidental pulmonary nodules. J Am Coll Radiol 2017;14:773-7. Crossref

9. Wiener RS, Gould MK, Slatore CG, Fincke BG, Schwartz LM, Woloshin S. Resource use and guideline concordance in evaluation of pulmonary nodules for cancer: too much and too little care. JAMA Intern Med 2014;174:871-80. Crossref

10. LeMense GP, Waller EA, Campbell C, Bowen T. Development and outcomes of a comprehensive multidisciplinary incidental lung nodule and lung cancer screening program. BMC Pulm Med 2020;20:115. Crossref

11. MacMahon H, Naidich DP, Goo JM, et al. Guidelines for management of incidental pulmonary nodules detected on CT images: from the Fleischner Society 2017. Radiology 2017;284:228-43. Crossref

12. Bai C, Choi CM, Chu CM, et al. Evaluation of pulmonary nodules: clinical practice consensus guidelines for Asia. Chest 2016;150:877-93. Crossref

13. Detterbeck FC, Boffa DJ, Kim AW, Tanoue LT. The eighth edition lung cancer stage classification. Chest 2017;151:193-203. Crossref

14. Bach PB, Mirkin JN, Oliver TK, et al. Benefits and harms of CT screening for lung cancer: a systematic review. JAMA 2012;307:2418-29. Crossref

15. Field JK, Duffy SW, Baldwin DR, et al. UK Lung Cancer RCT Pilot Screening Trial: baseline findings from the screening arm provide evidence for the potential implementation of lung cancer screening. Thorax 2016;71:161-70. Crossref

16. Pedersen JH, Ashraf H, Dirksen A, et al. The Danish randomized lung cancer CT screening trial—overall design and results of the prevalence round. J Thorac Oncol 2009;4:608-14. Crossref

17. Horeweg N, Scholten ET, de Jong PA, et al. Detection of lung cancer through low-dose CT screening (NELSON): a prespecified analysis of screening test performance and interval cancers. Lancet Oncol 2014;15:1342-50. Crossref

18. Krist AH, Davidson KW, Mangione CM, et al. Screening for lung cancer: US Preventive Services Task Force recommendation statement. JAMA 2021;325:962-70. Crossref

19. Weinstock TG, Tewari A, Patel H, et al. No stone unturned: Nodule Net, an intervention to reduce loss to follow-up of lung nodules. Respir Med 2019;157:49-51. Crossref

20. Chang GC, Chiu CH, Yu CJ, et al. Low-dose CT screening among never-smokers with or without a family history of lung cancer in Taiwan: a prospective cohort study. Lancet Respir Med 2024;12:141-52. Crossref

21. ClinicalTrials.gov. Screening for lung cancer in subjects with family history of lung cancer (Identifier NCT05762731). 2023. Available from: https://www.clinicaltrials.gov/study/NCT05762731 . Accessed 19 Apr 2024.

22. Schmid-Bindert G, Vogel-Claussen J, Gütz S, et al. Incidental pulmonary nodules—what do we know in 2022. Respiration 2022;101:1024-34. Crossref

23. Digby GC, Habert J, Sahota J, Zhu L, Manos D. Incidental pulmonary nodule management in Canada: exploring current state through a narrative literature review and expert interviews. J Thorac Dis 2024;16:1537-51. Crossref

24. Carr LL, Dyer DS, Zelarney PT, Kern EO. Improvement in stage of lung cancer diagnosis with incidental pulmonary nodules followed with a patient tracking system and computerized registry. JTO Clin Res Rep 2022;3:100297. Crossref

25. Dyer DS, Zelarney PT, Carr LL, Kern EO. Improvement in follow-up imaging with a patient tracking system and computerized registry for lung nodule management. J Am Coll Radiol 2021;18:937-46. Crossref

26. O’Keeffe LM, Taylor G, Huxley RR, Mitchell P, Woodward M, Peters SA. Smoking as a risk factor for lung cancer in women and men: a systematic review and meta-analysis. BMJ Open 2018;8:e021611. Crossref

27. McCarthy WJ, Meza R, Jeon J, Moolgavkar SH. Chapter 6: Lung cancer in never smokers: epidemiology and risk prediction models. Risk Anal 2012;32 Suppl 1(Suppl 1):S69-84. Crossref

28. Church TR, Black WC, Aberle DR, et al. Results of initial low-dose computed tomographic screening for lung cancer. N Engl J Med 2013;368:1980-91. Crossref

29. Kovalchik SA, Tammemagi M, Berg CD, et al. Targeting of low-dose CT screening according to the risk of lung-cancer death. N Engl J Med 2013;369:245-54. Crossref

30. Aberle DR, DeMello S, Berg CD, et al. Results of the two incidence screenings in the National Lung Screening Trial. N Engl J Med 2013;369:920-31. Crossref

31. Patz EF Jr, Pinsky P, Gatsonis C, et al. Overdiagnosis in low-dose computed tomography screening for lung cancer. JAMA Intern Med 2014;174:269-74. Crossref

32. Hong Kong Cancer Registry, Hospital Authority. Cancer Facts. 2020. Available from: https://www3.ha.org.hk/cancereg/facts.html. Accessed 19 Apr 2024.

33. Cho J, Choi SM, Lee J, et al. Proportion and clinical features of never-smokers with non–small cell lung cancer. Chin J Cancer 2017;36:20.Crossref

34. Chinese Expert Group on Early Diagnosis and Treatment of Lung Cancer, China Lung Oncology Group. China National Lung Cancer Screening Guideline with Lowdose Computed Tomography (2023 Version) [in Chinese]. Zhongguo Fei Ai Za Zhi 2023;26:1-9. Crossref

35. MacRosty CR, Rivera MP. Lung cancer in women: a modern epidemic. Clin Chest Med 2020;41:53-65. Crossref

36. Horeweg N, van Rosmalen J, Heuvelmans MA, et al. Lung cancer probability in patients with CT-detected pulmonary nodules: a prespecified analysis of data from the NELSON trial of low-dose CT screening. Lancet Oncol 2014;15:1332-41. Crossref

37. Tolwin Y, Gillis R, Peled N. Gender and lung cancer–SEER-based analysis. Ann Epidemiol 2020;46:14-9. Crossref

38. Flores R, Patel P, Alpert N, Pyenson B, Taioli E. Association of stage shift and population mortality among patients with non–small cell lung cancer. JAMA Netw Open 2021;4:e2137508. Crossref

39. Potter AL, Rosenstein AL, Kiang MV, et al. Association of computed tomography screening with lung cancer stage shift and survival in the United States: quasi-experimental study. BMJ 2022;376:e069008. Crossref

Prostate cancer is one of the most common cancers in men. In 2020, there were 2315 new cases of prostate cancer diagnosed in Hong Kong, with an age-standardised incidence rate of 30.5 per 100 000 population.1 Globally, >1.4 million new prostate cancer cases and more than 370 000 related deaths were reported in 2020.2 Robot-assisted radical prostatectomy (RARP) is one of the most common procedures used to eradicate localised prostate cancer. However, erectile dysfunction (ED) and urinary incontinence (UI) are common side-effects of RARP.3

The RARP is typically performed using robotic surgical platforms, such as the da Vinci Surgical System4 which facilitates minimally invasive prostatectomy. Keyholes are created through which high-resolution, illuminated cameras and robotic arms are inserted into the peritoneal cavity, which is inflated with carbon dioxide to provide adequate space for surgery. If the tumour is small and likelihood of extracapsular extension (ECE) is low, bilateral or unilateral nerve sparing (NS) may be performed to preserve postoperative erectile5 and lower urinary tract function,6 7 while taking oncological outcomes into consideration. The decision to use an NS technique is made by the surgeon, who carefully assesses the patient’s disease characteristics, drawing on personal experience and current research evidence.8 The prostate is then dissected from the bladder and urethra, and a re-anastomosis is performed between the bladder neck and the urethra.

A meta-analysis of NS techniques in radical prostatectomy (including RARP) has shown that the use of NS techniques results in lower risks of ED and UI at 3- and 12-month follow-ups.9 Nerve sparing cases demonstrate superior erectile function, urinary continence, and oncological outcomes compared with non-NS cases.9 Further analyses indicated that NS is associated with fewer complications than non-NS.10 11 They also suggest that the use of NS techniques does not lead to inferior oncological outcomes.10 11 Therefore, we hypothesised that bilateral NS in RARP improves erectile function and urinary continence after surgery. Our study aimed to investigate the effects of NS RARP on the aforementioned side-effects of ED and UI in Hong Kong and provide suggestions for enhancing patient’s quality of life.

We retrospectively recruited 431 patients who underwent RARP in a university-based teaching hospital (our institution) between January 2018 and April 2023. We retrieved their basic demographics, relevant surgical parameters (NS approach, positive surgical margin [PSM], ECE and Gleason score), postoperative 1-hour pad test results (at 1, 2, 3, 6, and 12 months postoperatively), and pre- and postoperative International Index of Erectile Function–5 (IIEF-5) scores12 at the same time points from electronic medical records. The IIEF-5 assesses erectile function using a 5-point scale across several domains, including erectile function, orgasmic function, sexual desire, intercourse satisfaction, and overall sexual satisfaction.12 All patients attended follow-up at our institution’s urology nurse clinic and received guidance on postoperative management, including pelvic floor strengthening exercises.

The primary outcome of the study was to evaluate the effect of NS in RARP on postoperative ED and UI. Secondary outcomes included correlations between other factors (PSM and ECE) and functional outcomes (ED and UI); correlations between NS and PSM or ECE; and postoperative trends in ED and UI beyond 1 year.

Inclusion criteria comprised all patients who had undergone RARP in our institution with follow-up in our nurse clinic. Treatment via bilateral NS, unilateral NS, or non-NS RARP was performed at the surgeon’s discretion and the patient’s preferences. Exclusion criteria included incomplete data. For the ED outcome, patients with an IIEF-5 score of 7 or below were excluded because this score indicates severe ED,13 and improvement beyond the preoperative baseline was not expected after NS RARP. Patient selection flowcharts for ED and UI are provided in online supplementary Figures 1 and 2, respectively.

R (R Foundation for Statistical Computing, Vienna, Austria) and RStudio software were used for data analysis.14 All statistical tests were two-sided and incorporated a 5% significance threshold. Cases with missing data due to loss to follow-up were excluded from analysis. Patients selected for ED and UI analysis had comparable age profiles (online supplementary Tables 1 and 2, respectively). Categorical variables were analysed using Chi squared tests or Fisher’s exact tests, depending on the observed frequencies. Continuous variables were analysed using independent sample t tests and Pearson correlations were utilised.

Among the 431 eligible patients included in the analysis, the mean age was 67.67 years. Regarding ED, the mean ages (±standard deviation) in the non-NS, unilateral NS, and bilateral NS groups were 65.20±4.95, 64.84±6.51, and 66.64±5.35 years, respectively, with no statistically significant differences observed (online supplementary Table 3). Concerning UI, the mean ages (±standard deviation) in the non-NS, unilateral NS, and bilateral NS groups were 66.82±5.48, 67.90±5.49, and 67.94±5.14 years, respectively, with no statistically significant differences observed (online supplementary Table 4). The mean tumour percentage was 12.85% of the total prostate volume and the mean resected prostate volume was 54.39 g. The distribution of pathological Gleason scores is shown in online supplementary Table 5. The majority of patients had a Gleason score of 7: 41.0% had a score of 3+4 (International Society of Urological Pathology [ISUP] Grade Group 2), and 23.9% had a score of 4+3 (ISUP Grade Group 3). Patients with low to intermediate risk prostate cancer, based on the National Comprehensive Cancer Network risk classification, were more likely to undergo bilateral NS operations (Table 1). Further details of NS approaches are presented in Table 1.

The mean preoperative IIEF-5 score was 10.22 (n=75; online supplementary Fig 1). For the primary outcome, patients with bilateral NS had a higher mean postoperative IIEF-5 score than those without NS at 2 months (non-NS vs bilateral NS=3.19 vs 7.60, t=-2.35; P=0.037). Bilateral NS patients also had a higher mean postoperative IIEF-5 score than unilateral NS patients at 2 months (2.50 vs 7.60, t=-2.69; P=0.020) and 3 months (2.06 vs 7.40, t=-2.61; P=0.027) [Table 2]. Differences in IIEF-5 scores at 1, 6, and 12 months postoperatively were not significant among any of the groups (Table 2). Concerning the secondary outcome, younger age was associated with a higher postoperative IIEF-5 score among non-NS patients (m=-0.42; P=0.024).

With respect to postoperative penile rehabilitation, the use of phosphodiesterase type 5 inhibitors (PDE5i) was reported by 38.3%, 77.8% and 60.0% in non-NS, unilateral NS, and bilateral NS patients, respectively. Among those who did not take PDE5i, reasons included financial considerations—patients must pay out of pocket for PDE5i in Hong Kong, and a perceived lack of efficacy. Even among those who utilised PDE5i, only 38.9%, 35.7%, and 66.7% of non-NS, unilateral NS, and bilateral NS patients, respectively, reported subjective improvement in erectile function. Objective changes in IIEF-5 scores (difference between preoperative and 12-month postoperative scores) were -10.20, -17.03, and -7.67 in non-NS, unilateral NS, and bilateral NS patients, respectively. Further details can be found in online supplementary Tables 6 to 8.

After initial screening and the exclusion of records with missing follow-up data, 264 patients were included in the analysis of UI following RARP (online supplementary Fig 2).

For the primary outcome, patients with bilateral NS had lower mean urinary leakage volume in the 1-hour pad test relative to patients without NS after 1 month (non-NS vs bilateral NS=49.44 g vs 16.40 g, t=3.92; P<0.001) and 2 months (35.45 g vs 13.60 g, t=2.67; P=0.009). Patients with bilateral NS also had lower mean urinary leakage volume than unilateral NS patients after 1 month (50.82 g vs 16.40 g, t=2.61; P=0.01). Differences in UI at 3, 6, and 12 months were not significant among the groups (Table 3). Further details comparing non-NS and unilateral NS can be found in online supplementary Tables 9 and 10.

Tumour recurrence is an adverse surgical outcome of RARP that requires further oncological management. We identified patients who underwent adjuvant radiotherapy, salvage radiotherapy, or experienced cancer-related death, then stratified them according to NS group, adjusted for age and total Gleason score. Statistical analysis showed no significant differences in oncological outcomes between non-NS and bilateral NS patients (odds ratio [OR]=0.75, 95% confidence interval [95% CI]=0.39-1.27; P=0.321), as well as unilateral NS and bilateral NS patients (OR=0.78, 95% CI=0.18-2.82; P=0.720). These findings indicated that bilateral NS was neither superior nor inferior in oncological outcomes compared with unilateral and non-NS groups, consistent with literature reports.15 16

Older patients had lower postoperative IIEF-5 scores at 6 months (Pearson correlation=-0.18; P=0.013) and 12 months (Pearson correlation=-0.22; P=0.014). However, the correlations were not statistically significant at 1, 2, or 3 months postoperatively. There also was no statistically significant correlation between age and postoperative UI.

Patients who underwent non-NS RARP were more likely to have ≥1 positive surgical margin (Chi squared=4.2673, P=0.039, OR=0.46; 95% CI=0.22-0.91). This result was likely attributable to disease-related factors. A larger proportion of patients in the non-NS group had a higher National Comprehensive Cancer Network risk score, indicating more aggressive tumours. Distributions of patients’ NS status, tumour grading, and PSMs are shown in Table 1.

Comparisons of postoperative IIEF-5 scores among NS groups revealed significant differences. Patients with bilateral NS exhibited higher postoperative IIEF-5 scores than those without NS at 2 months, highlighting the positive impact of bilateral NS on early erectile function recovery. Similar trends were observed when comparing bilateral NS with unilateral NS at both 2 and 3 months postoperatively. However, no significant differences in IIEF-5 scores were noted at 6 or 12 months, suggesting a convergence of outcomes beyond the initial recovery phase. A meta-analysis Nguyen et al17 on NS techniques in radical prostatectomy (including RARP) showed that NS cases had lower risks of ED at 3 and 12 months (risk ratio [RR] at 3 months=0.77; 95% CI=0.70-0.85; RR at 12 months=0.53; 95% CI=0.39-0.71). Some differences between our study (Table 2) and that of Nguyen et al17 might be attributable to the small sample size and loss to follow-up in our cohort, which may have introduced selection bias.

Our study also demonstrated that the bilateral NS technique was linked to better erectile function outcomes than unilateral NS, consistent with previous findings. In the study by Berg et al,18 the proportion of patients who were alive, continent, and potent was significantly greater among those with bilateral NS (67.6%) compared to unilateral NS (31.3%) [P<0.001]. Other studies also showed that bilateral NS was associated with lower risks of ED at 3 months (RR=0.80; 95% CI=0.70-0.90) and 1 year (RR=0.80; 95% CI=0.72-0.88) relative to unilateral NS.17

However, our study did not identify statistically significant benefits from unilateral NS in terms of ED and UI compared with non-NS. This may be due to a combination of factors, including variation in surgeons’ techniques for unilateral NS, small sample size, and relatively small absolute differences in outcomes between the unilateral and non-NS groups.

Further analysis also found that PDE5i therapy yielded only modest improvements in erectile function among individuals who underwent non-NS or unilateral NS procedures. Changes in IIEF-5 scores from preoperative to 12 months postoperative after PDE5i use were -10.20, -17.03 and -7.67 in non-NS, unilateral NS, and bilateral NS patients, respectively. Although approximately two-thirds of bilateral NS patients experienced improved erectile function, the limited number of patients who received PDE5i warrants caution when interpreting effects in this subgroup.

The relationship between age and erectile function in patients without NS was notable. Younger age was associated with higher postoperative IIEF-5 scores, emphasising the potential influence of age on postoperative erectile function recovery, consistent with international literature.19 Among older patients, postoperative ED remains an important complication. Thus, age is a key consideration when balancing quality of life and oncological control in planning surgical approaches. Surgeons should thoroughly discuss potential side-effects of ED with older patients who have concerns about sexual function prior to surgery involving NS.

In terms of UI, patients with bilateral NS demonstrated lower urinary leakage volumes in the 1-hour pad test compared with those without NS at 1 and 2 months postoperatively. Similarly, bilateral NS was associated with lower urinary leakage volumes than unilateral NS at 1 month. These findings suggest that NS techniques enhance short-term UI recovery, although the differences tend to diminish over time. The benefit of NS on long-term UI recovery beyond 1 year remains uncertain.9 17 20 Choi et al20 reported findings similar to ours, indicating that bilateral NS was associated with a higher continence rate than non-NS at 4 months (P=0.043), although the differences at 12 and 24 months were not significant. Conversely, a meta-analysis by Nguyen et al17 on NS techniques in radical prostatectomy (including RARP) found that NS was associated with lower risks of UI at both 3 months (RR =0.75, 95% CI=0.65-0.85) and 12 months (RR=0.61, 95% CI=0.44-0.84). These discrepancies may be attributed to differences in study design, sample size, and baseline patient characteristics.

Other studies have also shown that bilateral NS is associated with a lower risk of UI at 1 year (RR=0.70, 95% CI=0.50-0.98) and lower risks of ED at 3 months (RR=0.80, 95% CI=0.70-0.90) and 1 year (RR=0.80, 95% CI=0.72-0.88) compared with unilateral NS.17 In our study, while we observed a lower risk of UI at 1 and 2 months after surgery, this difference was not statistically significant at 3 months.

Our findings have implications for surgical decision-making and patient counselling. Surgeons should consider the potential benefits of bilateral NS for early postoperative recovery of erectile function and urinary continence.21 Although our study did not demonstrate statistically significant long-term differences in these outcomes, bilateral NS has been reported to play a key role in early recovery. Contributing to improved quality of life and patient satisfaction.21 Moreover, age should be considered when evaluating erectile function outcomes in patients undergoing non-NS RARP.

Future research could investigate the long-term trajectories of erectile function and urinary continence, exploring factors that contribute to outcome convergence over time. Additionally, efforts to evaluate the impact of NS techniques on quality of life and patient satisfaction could offer a more comprehensive understanding of the clinical implications of these findings.

We also intend to further examine why patients with unilateral NS did not demonstrate better erectile function outcomes than those in the non-NS group. Notably, there was a significant difference in outcomes between the bilateral NS and non-NS groups, despite the small number of patients with bilateral NS (n=10). Therefore, we plan to conduct a detailed review of surgical records for patients with unilateral NS to determine why erectile function outcomes were not superior to those in the non-NS group.

This study also highlights the need for better public and patient education regarding sexual health. In Hong Kong, sexual function currently remains a major taboo topic among older individuals. Many are reluctant to discuss ED and are even more reluctant to seek medical treatment. In our study, 121 patients displayed severe preoperative ED, but none had sought medical attention. This problem is compounded by the financial barriers to treatment. In public hospitals, PDE5i is entirely self-financed, and even government employees are unable to reclaim their costs. These factors collectively create a substantial barrier for older men in Hong Kong to recognise ED as a treatable clinical condition that could improve their sexual health.

To our knowledge, this is the first local retrospective cohort study in Hong Kong comparing the efficacy of NS approaches in RARP for reducing ED and UI. These findings provide valuable insights into current local RARP practices and serve as a foundation for future prospective studies.

However, there were several limitations, including the retrospective design, potential selection biases, single-centre setting, and inconsistency in follow-up intervals, as some patients might have attended follow-up appointments earlier or later than scheduled due to personal reasons. These factors could affect the interpretation and generalisability of the results. Furthermore, the IIEF-5 score used in this study was based on patients’ subjective self-assessment and may not accurately reflect changes in erectile function, particularly among patients without preoperative sexual activity.22

The bilateral NS technique during prostatectomy demonstrated a significant positive impact on the recovery of erectile function and urinary continence within the first 6 months postoperatively, without compromising oncological outcomes. However, the extent of this benefit appears to diminish over time, indicating the need for longer-term assessment. These findings contribute valuable insights into the role of NS in prostate cancer surgery and may inform clinical decision-making in prostate cancer management. To validate and expand upon these observations, further prospective, randomised studies with extended follow-up are warranted.

Concept or design: OWK Tsui, S Chun, BSH Ho.
Acquisition of data: OWK Tsui, KCH Shing.
Analysis or interpretation of data: OWK Tsui, KCH Shing, APM Lam, S Chun.
Drafting of the manuscript: OWK Tsui.
Critical revision of the manuscript for important intellectual content: All authors.

All authors had full access to the data, contributed to the study, approved the final version for publication, and take responsibility for its accuracy and integrity.

As an editor of the journal, BSH Ho was not involved in the peer review process. Other authors have disclosed no conflicts of interest.

This research was presented at The Hong Kong Urological Association 28th Annual Scientific Meeting held in Hong Kong on 19 November 2023.

This research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors.

This research was approved by the Institutional Review Board of The University of Hong Kong/Hospital Authority Hong Kong West Cluster, Hong Kong (Ref No.: UW-24-099). The requirement for informed patient consent was waived by the Board due to the retrospective nature of the research.

The supplementary material was provided by the authors and some information may not have been peer reviewed. Accepted supplementary material will be published as submitted by the authors, without any editing or formatting. Any opinions or recommendations discussed are solely those of the author(s) and are not endorsed by the Hong Kong Academy of Medicine and the Hong Kong Medical Association. The Hong Kong Academy of Medicine and the Hong Kong Medical Association disclaim all liability and responsibility arising from any reliance placed on the content.

1. Centre for Health Protection, Department of Health, Hong Kong SAR Government. Prostate cancer. 2025 Feb 24. Available from: https://www.chp.gov.hk/en/healthtopics/content/25/5781.html#:~:text=Prostate%20cancer%20was%20the%20fourth,per%20100%20000%20male%20population. Accessed 24 Feb 2025.

2. Wang L, Lu B, He M, Wang Y, Wang Z, Du L. Prostate cancer incidence and mortality: global status and temporal trends in 89 countries from 2000 to 2019. Front Public Health 2022;10:811044. Crossref

3. Kesch C, Heidegger I, Kasivisvanathan V, et al. Radical prostatectomy: sequelae in the course of time. Front Surg 2021;8:684088. Crossref

4. Deligiannis D, Anastasiou I, Mygdalis V, Fragkiadis E, Stravodimos K. Change of practice patterns in urology with the introduction of the Da Vinci surgical system: the Greek NHS experience in debt crisis era. Arch Ital Urol Androl 2015;87:56-61. Crossref

5. Harris CR, Punnen S, Carroll PR. Men with low preoperative sexual function may benefit from nerve sparing radical prostatectomy. J Urol 2013;190:981-6. Crossref

6. Kübler HR, Tseng TY, Sun L, Vieweg J, Harris MJ, Dahm P. Impact of nerve sparing technique on patient self-assessed outcomes after radical perineal prostatectomy. J Urol 2007;178:488-92; discussion 492. Crossref

7. Avulova S, Zhao Z, Lee D, et al. The effect of nerve sparing status on sexual and urinary function: 3-year results from the CEASAR study. J Urol 2018;199:1202-9. Crossref

8. Vis AN, van den Bergh RC, van der Poel HG, et al. Selection of patients for nerve sparing surgery in robot-assisted radical prostatectomy. BJUI Compass 2021;3:6-18. Crossref

9. Liu Y, Deng XZ, Qin J, et al. Erectile function, urinary continence and oncologic outcomes of neurovascular bundle sparing robot-assisted radical prostatectomy for high-risk prostate cancer: a systematic review and meta-analysis. Front Oncol 2023;13:1161544. Crossref

10. Furrer MA, Sathianathen N, Gahl B, et al. Oncological outcomes after attempted nerve-sparing radical prostatectomy (NSRP) in patients with high-risk prostate cancer are comparable to standard non-NSRP: a longitudinal long-term propensity-matched single-centre study. BJU Int 2024;133:53-62. Crossref

11. Takahara K, Sumitomo M, Fukaya K, et al. Clinical and oncological outcomes of robot-assisted radical prostatectomy with nerve sparing vs. non-nerve sparing for high-risk prostate cancer cases. Oncol Lett 2019;18:3896-902. Crossref

12. Díaz-Mohedo E, Meldaña Sánchez A, Cabello Santamaría F, Molina García E, Hernández Hernández S, Hita-Contreras F. The Spanish version of the International Index of Erectile Function: adaptation and validation. Int J Environ Res Public Health 2023;20:1830. Crossref

13. Yin Y, Wang K, Xu Y, et al. The impact of using donor sperm after ICSI failure in severe oligozoospermia on male mental health and erectile function. J Multidiscip Healthc 2024;17:21-8. Crossref

14. Gasparini A. comoRbidity: an R package for computing comorbidity scores. J Open Source Softw 2018;3:648. Crossref

15. Kumar A, Samavedi S, Bates AS, et al. Safety of selective nerve sparing in high-risk prostate cancer during robotassisted radical prostatectomy. J Robot Surg 2017;11:129-38. Crossref

16. Yuh B, Artibani W, Heidenreich A, et al. The role of robot-assisted radical prostatectomy and pelvic lymph node dissection in the management of high-risk prostate cancer: a systematic review. Eur Urol 2014;65:918-27. Crossref

17. Nguyen LN, Head L, Witiuk K, et al. The risks and benefits of cavernous neurovascular bundle sparing during radical prostatectomy: a systematic review and meta-analysis. J Urol 2017;198:760-9. Crossref

18. Berg KD, Thomsen FB, Hvarness H, Christensen IJ, Iversen P. Early biochemical recurrence, urinary continence and potency outcomes following robot-assisted radical prostatectomy. Scand J Urol 2014;48:356-66. Crossref

19. Tal R, Alphs HH, Krebs P, Nelson CJ, Mulhall JP. Erectile function recovery rate after radical prostatectomy: a metaanalysis. J Sex Med 2009;6:2538-46. Crossref

20. Choi WW, Freire MP, Soukup JR, et al. Nerve-sparing technique and urinary control after robot-assisted laparoscopic prostatectomy. World J Urol 2011;29:21-7. Crossref

21. Ngoo KS, Honda M, Kimura Y, et al. Longitudinal study on the impact of urinary continence and sexual function on health-related quality of life among Japanese men after robot-assisted radical prostatectomy. Int J Med Robot 2019;15:e2018. Crossref

22. Tang Z, Li D, Zhang X, et al. Comparison of the simplified International Index of Erectile Function (IIEF-5) in patients of erectile dysfunction with different pathophysiologies. BMC Urol 2014;14:52. Crossref