Stevens–Johnson syndrome (SJS) and toxic epidermal necrolysis (TEN) are uncommon but potentially life-threatening severe mucocutaneous reactions characterised by extensive epidermal necrosis and detachment. Both entities are considered variants of a single disease continuum and are classified according to the percentage of body surface area (BSA) with epidermal detachment.1 2 Although SJS and TEN (hereafter, SJS/TEN) have been described in all ethnicities worldwide,3 studies of these reactions in Hong Kong have been limited.4 5 The incidence, clinical characteristics, aetiology, treatment regimen, morbidity, and mortality in the territory are largely unknown. This pilot study aimed to review cases of SJS/TEN over a 17-year period at a tertiary referral centre in Hong Kong, and to aid future research in Hong Kong focused on severe cutaneous adverse reactions.

This retrospective cohort study included all hospitalised patients who had been diagnosed with SJS/TEN and were treated in Prince of Wales Hospital (PWH), a major regional hospital under the New Territories East Cluster (NTEC), from 1 January 2004 to 31 December 2020.

Patients with clinical and histological diagnoses of SJS/TEN were identified from the Hospital Authority database using International Classification of Diseases, Ninth Revision, Clinical Modification diagnosis codes and the database of the Department of Anatomical and Cellular Pathology of PWH, respectively.

Diagnoses of SJS/TEN were based on consensus guidelines.1 Patients were diagnosed with SJS, SJS/TEN overlap, and TEN when they exhibited epidermal detachment levels of <10%, 10% to 30%, and >30%, respectively, with consistent histological features (if skin biopsy was performed). Consistent histological features were regarded as partial- to full-thickness epidermal necrosis.

Patients were excluded if they had an alternative diagnosis, such as severe cutaneous adverse reactions other than SJS/TEN (eg, drug reaction with eosinophilia and systemic symptoms syndrome/acute generalised exanthematous pustulosis/generalised bullous fixed drug eruption), erythema multiforme major, autoimmune blistering disease, acute graft-versus-host disease, and infections such as staphylococcal scalded skin syndrome.

Clinical characteristics were collected from electronic records and, when available, hospital case notes. The following clinical characteristics were recorded and analysed: age at onset, sex, ethnicity, maximal BSA of detached or detachable skin, SCORTEN (Severity-of-Illness Score for Toxic Epidermal Necrolysis) prognostic score,6 mucosa involved, histology results if available, causative drugs, time from exposure to onset, time from onset to admission and treatment, treatment regimen, disease complications, mortality and its cause, and length of stay. Efforts to identify causative drugs were guided by the ALDEN (algorithm of drug causality for epidermal necrolysis) score,7 which was retrospectively calculated by two independent investigators. All clinical data were expressed as percentages or means ± standard deviations unless otherwise specified.

This article was written in compliance with the STROBE (Strengthening the Reporting of Observational Studies in Epidemiology) reporting guidelines.

Using the International Classification of Diseases, Ninth Revision, Clinical Modification diagnosis codes for SJS/TEN, 164 potential patients with 166 cases of SJS/TEN during the period from January 2004 to December 2020 were initially identified. Six additional patients with SJS/TEN were identified from the database of the Department of Anatomical and Cellular Pathology of PWH. Forty-seven patients were excluded due to alternative diagnoses. In total, 123 patients with 125 cases of SJS/TEN were included in the study (Fig).

Among the 123 patients with SJS/TEN, 53 were men and 70 were women; the female-to-male ratio was 1.32:1 (Table 1). The mean age at onset was 51.4 years, and most patients were Chinese. Of the 125 cases, 59 were SJS, 27 were SJS-TEN overlap, and 39 were TEN. A small number of patients (n=18, 14.4%) were admitted for other medical issues but developed SJS/TEN after hospitalisation.

The mean time from disease onset to hospitalisation was 4.9 days (Table 2). Fever was present on admission in 88 cases (70.4%). The mean maximal BSA of epidermal detachment was 23%. Mucosal involvement was common; only five cases (4.0%) lacked mucosal involvement. Skin biopsy was performed in 87 cases (69.6%) and the mean SCORTEN prognostic score was 2.17.

Burns unit or intensive care unit admission was required in 40 cases (32.0%) and half of these cases required invasive mechanical ventilation. In total, 63 cases (50.4%) involved concomitant infection from various sources. Multiorgan failure or disseminated intravascular coagulation occurred in 29 cases (23.2%). The mean length of stay in the hospital was 23.9 days (Table 2).

Stevens–Johnson syndrome and TEN onset was attributed to a drug in 114 of 124 cases (91.9%); one patient developed a second case of SJS/TEN upon accidental re-exposure to the same culprit drug (paracetamol). Four cases (3.2%) were caused by infection, and no cause was identified in six cases (4.8%). The identified culprit drugs are shown in Table 3. The mean time from initiation of the culprit drug to onset of SJS/TEN was 20.5 ± 16.7 days (range, 1-87; median, 15.5).

All patients received the best supportive medical care available. In some cases, patients received additional treatment. The numbers and proportions of cases treated with different regimens are shown in Table 4. The mean time between disease onset and active treatment initiation was 7.4 ± 6.1 days (range, 1-34; median, 6).

Intravenous immunoglobulin (IVIG) was administered in 54 cases. The mean total dose of IVIG was 3.2 g/kg (range, 1.5-6; administered over 2-6 days). High-dose IVIG, defined as ≥3 g/kg, was administered in 40 cases. Systemic steroid regimens considerably varied, with daily doses of prednisolone ranging from 20 to 120 mg (or an equivalent dose). The cyclosporine regimen was 3 mg/kg/day, tapered over 20 to 30 days.

There were 27 deaths in the study cohort, and the overall mortality rate was 21.6%. The mean time from SJS/TEN onset to death was 23.6 ± 19.0 days (range, 5-76). Most patients died from multiorgan failure and/or fulminant sepsis (n=22, 81.5%); other causes of death were acute coronary syndrome (n=2), liver failure (n=1), and sudden cardiac arrest (n=2).

The observed mortality rates were 16.9%, 22.2%, and 28.2% for SJS, SJS-TEN overlap, and TEN, respectively. The SCORTEN-based predicted mortality rates were 14.1%, 28.5%, and 36.9% for SJS, SJS-TEN overlap, and TEN, respectively. Inpatient-onset SJS/TEN had a high mortality rate of 77.8%: 14 deaths among 18 patients who developed SJS/TEN after admission.

Stevens–Johnson syndrome and TEN are recognised worldwide, with several epidemiological studies conducted in Europe and the US. In the 1990s, Roujeau et al8 reported that the annual incidence of TEN in France was 1.2 cases per million; during the same period, the estimated overall annual incidences of SJS/TEN were 1.89 cases per million in Germany9 and 4.2 cases per million in the US.10 In the past decade, two large epidemiological studies in the US11 and United Kingdom12 revealed that the overall annual incidences of SJS/TEN were 12.7 and 5.76 cases per million, respectively. In contrast, the epidemiology of SJS/TEN in Asia is not well-documented.13 In Singapore, based on a small retrospective hospital-based study of 20 patients with TEN, the estimated annual incidence of TEN was 1.4 cases per million14; in Korea, a large population-based study indicated that the overall annual incidence of SJS/TEN was 4.9 to 6.5 cases per million.15 In the present study, there were 125 cases of SJS/TEN during the 17-year study period. Notably, 13 of these cases were transferred from hospitals outside of the NTEC: one was SJS, three were SJS/TEN overlap, and nine were TEN. The NTEC serves a population of 1.3 million.16 The estimated annual incidence of TEN alone and combined annual incidence of SJS, SJS-TEN overlap, and TEN were 1.36 and 5.07 cases per million, respectively; these incidences are comparable with findings from studies in other countries.

Stevens–Johnson syndrome is approximately threefold more common than TEN.15 17 However, in the current study, fewer than half of the cases (47.2%) were SJS, whereas 31.2% were TEN. This may be related to referral bias, whereby more severe cases were transferred to the study hospitals, whereas ‘milder’ cases were managed in regional hospitals where the patients were initially admitted. Prince of Wales Hospital is a tertiary referral centre and one of the few hospitals in Hong Kong with both on-site dermatologists and burns unit support. In our cohort, 31 cases (24.8%) were transferred from peripheral hospitals: 18 (14.4%) arrived from hospitals within the NTEC, whereas 13 (10.4%) arrived from hospitals outside of the NTEC.

Stevens–Johnson syndrome and TEN are most often drug-induced, and a culprit drug is identified in approximately 85% of cases.7 18 The reactions usually occur between 7 days and 8 weeks after drug ingestion.19 However, upon rechallenge with the culprit drug, SJS/TEN can develop within hours.17 19 Efforts to identify causative drugs were guided by the ALDEN score.7 In cases of SJS/TEN, the most commonly implicated high-risk medications are anticonvulsants, allopurinol, antimicrobials, and oxicam non-steroidal anti-inflammatory drugs.19 20 In the present study, SJS/TEN onset was attributed to a drug in 114 of 124 cases (91.9%). The mean time between drug initiation and SJS/TEN onset was 20.5 days. Among these 114 cases, 81.6% were caused by the high-risk medications listed above. These findings are comparable with previous reports.

Stevens–Johnson syndrome and TEN are associated with high mortality rates, with 1% to 5% in cases of SJS and 25% to 30% in cases of TEN. Survival analyses in multinational European studies (EuroSCAR [European Study of Severe Cutaneous Adverse Reactions] and RegiSCAR [Registry of Severe Cutaneous Adverse Reactions]) have indicated that the overall mortality rate in cases of SJS/TEN is approximately 22% to 23%.18 20 21 22 23 In Asia, reported overall mortality rates vary from 12.3% to 25%.24 25 26 27 Sepsis leading to multiorgan failure is the most common cause of death.21 Despite the substantial mortality, there currently is no therapeutic regimen with a clear benefit for patients with SJS/TEN.18 21 Considering the rarity of these diseases, it is difficult to conduct randomised trials. Early recognition, rapid withdrawal of offending agents, and best supportive care remain the primary components of clinical management.

In the current study, the overall mortality rate was 21.6%; in 81.5% of these cases, the patient died of fulminant sepsis or multiorgan failure. These findings are consistent with existing literature. However, the mortality rate in cases of SJS was much higher in the present study than in previous studies. In the 59 cases of SJS, there were 10 deaths; the mortality rate was 16.9%. Among the 10 patients who died, six experienced complete skin re-epithelisation before death from other medical conditions, which include massive duodenal ulcer bleeding, acute coronary syndrome, metastatic lung cancer, acute liver and renal failure due to herbs, aspiration pneumonia, and sudden cardiac arrest. The remaining four patients had inpatient-onset SJS; they were initially admitted for traumatic intracranial haemorrhage, post-hepatectomy liver failure, convulsions caused by metastatic lung cancer, and post-stroke seizure, respectively. These patients exhibited skin-specific improvements but soon died of aspiration pneumonia and acute renal failure, liver failure, metastatic lung cancer with respiratory failure and liver failure, and sudden cardiac arrest, respectively. The high mortality rate among patients with SJS in the present study could be related to referral bias (as noted in the Epidemiology subsection above); specifically, more severe cases of SJS with co-morbidities and/or complications may have been transferred to our tertiary centre for medical care, whereas less severe cases of SJS might have been managed in regional hospitals where the patients were initially admitted. Indeed, the predicted mortality rate (based on the SCORTEN prognostic score) among cases of SJS in our cohort was 14%; this rate was similar to the observed mortality rate.

In the present study, inpatient-onset SJS/TEN had a high mortality rate (77.8%). Although high mortality of inpatient-onset SJS/TEN was not previously described in the literature, we speculate that this high mortality was related to the underlying medical conditions for which patients were initially admitted. The clinical characteristics of the 14 patients who died are presented in Table 5.

In addition to high mortality, SJS/TEN were associated with high rates of burns unit/intensive care unit admission (32%) and prolonged length of stay (mean=23.9 days) [Table 2], placing a considerable burden on the public healthcare system.

As a retrospective cohort study, the present study had some intrinsic limitations. Some hospital case notes (ie, from earlier in the study period) were no longer retrievable. Clinical characteristics such as the exact date of disease onset, precise total BSA involved, and detailed drug history (including over-the-counter medications/medications prescribed by private doctors) might not have been available for some of these older cases. Skin biopsies were performed in 70% of cases and might have been omitted in cases of terminal illness. Many patients with milder cases were lost to follow-up after discharge; thus, long-term sequelae were not well-documented.

Additionally, referral bias may have been present because PWH is a tertiary referral centre. Such bias could have led to underestimation of the true incidence of SJS and overestimation of the incidence of TEN; milder cases of SJS might have been managed in regional hospitals, whereas more severe cases of TEN were transferred to our centre for better care. Similarly, there may have been overestimation of various outcome measures including length of stay, complications, and mortality.

Nonetheless, this study had several strengths. To our knowledge, this is one of the largest single-centre studies regarding SJS/TEN in Asia; it included a homogenous group of predominantly Chinese patients. The patients were managed by the same dermatology team with a consistent diagnostic and therapeutic approach throughout the study period. Data collection was adequate, and exhaustive evaluation of drug history was feasible for cases with access to both electronic records and hospital case notes. To ensure accurate identification of causative drugs, the ALDEN score was retrospectively evaluated by two independent dermatology doctors during the study.

This is the first large study in Hong Kong to provide data regarding the epidemiology, disease characteristics and clinical course, aetiology, treatment regimen, and mortality of SJS/TEN. Although uncommon, SJS/TEN is associated with substantial morbidity and mortality. Therefore, in addition to increasing awareness of SJS/TEN among patients and clinicians, efforts should be made to optimise inpatient care among public hospitals in Hong Kong by establishing dedicated multidisciplinary teams that are experienced in the management of SJS/TEN.

Concept or design: CMT Cheung, MM Chang.
Acquisition of data: All authors.
Analysis or interpretation of data: All authors.
Drafting of the manuscript: CMT Cheung, AWS Chan.
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.

All authors have disclosed no conflicts of interest.

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 Joint Chinese University of Hong Kong–New Territories East Cluster Clinical Research Ethics Committee (Ref No.: 2017.424). The requirement for informed consent was waived by the Committee due to the retrospective nature of the research.

1. Bastuji-Garin S, Rzany B, Stern RS, Shear NH, Naldi L, Roujeau JC. Clinical classification of cases of toxic epidermal necrolysis, Stevens–Johnson syndrome, and erythema multiforme. Arch Dermatol 1993;129:92-6. Crossref

2. Roujeau JC. Stevens–Johnson syndrome and toxic epidermal necrolysis are severity variants of the same disease which differs from erythema multiforme. J Dermatol 1997;24:726-9. Crossref

3. Roujeau JC, Chosidow O, Saiag P, Guillaume JC. Toxic epidermal necrolysis (Lyell syndrome). J Am Acad Dermatol 1990;23:1039-58. Crossref

4. Ying S, Ho W, Chan HH. Toxic epidermal necrolysis: 10 years experience of a burns centre in Hong Kong. Burns 2001;27:372-5. Crossref

5. Yeung CK. Intravenous immunoglobulin treatment for Stevens–Johnson syndrome and toxic epidermal necrolysis [dissertation]. Queen Mary Hospital, The University of Hong Kong; 2004.

6. Bastuji-Garin S, Fouchard N, Bertocchi M, Roujeau JC, Revuz J, Wolkenstein P. SCORTEN: a severity-of-illness score for toxic epidermal necrolysis. J Invest Dermatol 2000;115:149-53. Crossref

7. Sassolas B, Haddad C, Mockenhaupt M, et al. ALDEN, an algorithm for assessment of drug causality in Stevens–Johnson syndrome and toxic epidermal necrolysis: comparison with case-control analysis. Clin Pharmacol Ther 2010;88:60-8. Crossref

8. Roujeau JC, Guillaume JC, Fabre JP, Penso D, Fléchet ML, Girre JP. Toxic epidermal necrolysis (Lyell syndrome). Incidence and drug etiology in France, 1981-1985. Arch Dermatol 1990;126:37-42. Crossref

9. Rzany B, Mockenhaupt M, Baur S, et al. Epidemiology of erythema exsudativum multiforme majus, Stevens–Johnson syndrome, and toxic epidermal necrolysis in Germany (1990-1992): structure and results of a population-based registry. J Clin Epidemiol 1996;49:769-73. Crossref

10. Chan HL, Stern RS, Arndt KA, et al. The incidence of erythema multiforme, Stevens–Johnson syndrome, and toxic epidermal necrolysis. A population-based study with particular reference to reactions caused by drugs among outpatients. Arch Dermatol 1990;126:43-7. Crossref

11. Hsu DY, Brieva J, Silverberg NB, Silverberg JI. Morbidity and mortality of Stevens–Johnson syndrome and toxic epidermal necrolysis in United States adults. J Invest Dermatol 2016;136:138797. Crossref

12. Frey N, Jossi J, Bodmer M, et al. The epidemiology of Stevens–Johnson syndrome and toxic epidermal necrolysis in the UK. J Invest Dermatol 2017;137:1240-7. Crossref

13. Lee HY, Martanto W, Thirumoorthy T. Epidemiology of Stevens–Johnson syndrome and toxic epidermal necrolysis in Southeast Asia. Dermatologica Sinica 2013;31:217-20. Crossref

14. Chan HL. Toxic epidermal necrolysis in Singapore, 1989 through 1993: incidence and antecedent drug exposure. Arch Dermatol 1995;131:1212-3. Crossref

15. Yang MS, Lee JY, Kim J, et al. Incidence of Stevens–Johnson syndrome and toxic epidermal necrolysis: a nationwide population-based study using national health insurance database in Korea. PLoS One 2016;11:e0165933. Crossref

16. Hospital Authority, Hong Kong SAR Government. New Territories East Cluster biennial report 2018-2020. Available from: https://www3.ha.org.hk/ntec/clusterreport/clusterreport2018-20/index.html. Accessed 8 Feb 2024.

17. Schwartz RA, McDonough PH, Lee BW. Toxic epidermal necrolysis: Part I. Introduction, history, classification, clinical features, systemic manifestations, etiology, and immunopathogenesis. J Am Acad Dermatol 2013;69:173. e1-13; quiz 185-6. Crossref

18. Creamer D, Walsh SA, Dziewulski P, et al. UK guidelines for the management of Stevens–Johnson syndrome/toxic epidermal necrolysis in adults 2016. J Plast Reconstr Aesthet Surg 2016;69:e119-53. Crossref

19. Roujeau JC, Kelly JP, Naldi L, et al. Medication use and the risk of Stevens–Johnson syndrome or toxic epidermal necrolysis. N Engl J Med 1995;333:1600-7. Crossref

20. Mockenhaupt M, Viboud C, Dunant A, et al. Stevens–Johnson syndrome and toxic epidermal necrolysis: assessment of medication risks with emphasis on recently marketed drugs. The EuroSCAR study. J Invest Dermatol 2008;128:35-44. Crossref

21. Schwartz RA, McDonough PH, Lee BW. Toxic epidermal necrolysis: Part II. Prognosis, sequelae, diagnosis, differential diagnosis, prevention, and treatment. J Am Acad Dermatol 2013;69:187.e1-16; quiz 203-4. Crossref

22. Roujeau JC, Stern RS. Severe adverse cutaneous reactions to drugs. N Engl J Med 1994;331:1272-85. Crossref

23. Sekula P, Dunant A, Mockenhaupt M, et al. Comprehensive survival analysis of a cohort of patients with Stevens–Johnson syndrome and toxic epidermal necrolysis. J Invest Dermatol 2013;133:1197-204. Crossref

24. Barvaliya M, Sanmukhani J, Patel T, Paliwal N, Shah H, Tripathi C. Drug-induced Stevens–Johnson syndrome (SJS), toxic epidermal necrolysis (TEN), and SJS-TEN overlap: a multicentric retrospective study. J Postgrad Med 2011;57:115-9. Crossref

25. Roongpisuthipong W, Prompongsa S, Klangjareonchai T. Retrospective analysis of corticosteroid treatment in Stevens–Johnson syndrome and/or toxic epidermal necrolysis over a period of 10 years in Vajira Hospital, Navamindradhiraj University, Bangkok. Dermatol Res Pract 2014;2014:237821. Crossref

26. Suwarsa O, Yuwita W, Dharmadji HP, Sutedja E. Stevens–Johnson syndrome and toxic epidermal necrolysis in Dr Hasan Sadikin General Hospital Bandung, Indonesia from 2009-2013. Asia Pac Allergy 2016;6:43-7. Crossref

27. Lee HY, Fook-Chong S, Koh HY, Thirumoorthy T, Pang SM. Cyclosporine treatment for Stevens–Johnson syndrome/toxic epidermal necrolysis: retrospective analysis of a cohort treated in a specialized referral center. J Am Acad Dermatol 2017;76:106-13. Crossref

Intensive care treatments are primarily intended to improve patient outcomes. Considering the high operating costs of intensive care units (ICUs), a reliable, decision-supporting, risk stratification system is needed to predict patient outcomes and facilitate cost-effective use of ICU beds. Several disease severity scoring systems, such as the Acute Physiology and Chronic Health Evaluation (APACHE) system and the Simplified Acute Physiology Score system, are currently used to objectively assess outcomes and recovery potential in this complex and diverse group of patients.1 2

The APACHE system, one of the most commonly used benchmark severity scoring systems worldwide, can measure disease severity and predict hospital mortality among ICU patients. In the 40 years since its initial development, the APACHE system has undergone multiple revisions to improve statistical power and discrimination performance by modifying the numbers and weights of included variables.3 4 5 6 The underlying statistical principle is multivariable logistic regression based on data from an American population. The results are easy to interpret and allow robust outcome prediction for individuals with characteristics similar to the original population. However, the APACHE system has limited capacity to manage non-linear relationships between predictor and outcome variables, interactions between variables, and missing data. Although the value of the APACHE system for mortality prediction has been established, especially in Western countries, its discrimination performance and calibration are inconsistent when applied outside of the US.7 8 9 10 Since 2008, the Hospital Authority in Hong Kong has utilised the APACHE IV model to assess outcomes in critically ill patients. Nevertheless, the APACHE II model remains the most extensively validated version; it is widely used for research and reference purposes.11

In the early 1990s, artificial neural networks (ANNs), a type of machine learning algorithm, were proposed as alternative statistical techniques to logistic regression–based method. Similar to the organisation and data processing configurations in human brains, these networks consist of input and output layers with at least one or more intermediate (hidden) layers for pattern recognition. Each layer contains several ‘artificial neurons’, known as nodes, for data extraction; these nodes are connected with each other through variable ‘weights’.12 Artificial neural networks identify representative patterns from input data and observed output data within a training set, then fine-tune the variable weights; thus, they can predict outcomes when provided novel information. This method has considerable advantages in terms of managing non-linear relationships and multivariable interactions.13

A review of 28 studies comparing ANN and regression-based models showed that ANN outperformed regression-based models in 10 studies (36%), was outperformed by regression-based models in four studies (14%), and had similar performance in the remaining 14 studies (50%).14 Multiple recent studies also demonstrated that the integration of machine learning with electronic health records provided more accurate and reliable predictive performance compared with conventional prognostic models.15 16

This study was conducted to compare ANN performance with the performances of extensively validated and benchmark scoring systems—APACHE II and APACHE IV—in terms of predicting hospital mortality among critically ill patients in Hong Kong.

This retrospective analysis included all patients aged ≥18 years with first-time admissions to the ICU of Pamela Youde Nethersole Eastern Hospital between 1 January 2010 and 31 December 2019. The hospital is a 2000-bed tertiary care regional hospital that provides comprehensive services except for cardiothoracic surgery, transplant surgery, and burn management. The ICU is a 24-bed, closed, mixed medical-surgical unit with an average of 1600 patients admitted annually.

Demographic characteristics and hospital mortality data were retrospectively recorded. The worst value of each physiological parameter during the first 24 hours after ICU admission was used to generate an APACHE score. The predicted mortality risk was calculated based on published methods.3 5 Included parameters were age, sex, systolic and diastolic blood pressures, temperature, heart rate, respiratory rate, glucose level, blood urea nitrogen level, serum sodium level, creatinine level, haematocrit level, white cell count, albumin level, bilirubin level, pH, fraction of inspired oxygen, partial pressures of carbon dioxide and oxygen, bicarbonate, and urine output during the first 24 hours after ICU admission. For patients who had multiple ICU admissions during a single hospital stay, only the first admission was included. Patients were excluded if they died or were discharged from the ICU within 4 hours after admission.

Instances of incomplete data were resolved by multiple imputation using the Markov chain Monte Carlo algorithm (ie, fully conditional specification). This method fits a univariate (single dependent variable) model using all other available variables in the model as predictors, then imputes missing values for the dependent variable. The method continues until the maximum number of iterations is reached; the resulting imputed values are saved to the imputed dataset.

Neural network models were constructed with SPSS software (Windows version 25.0; IBM Corp, Armonk [NY], US) using the same parameters as in the APACHE IV model (online supplementary Fig); SPSS software was also used to examine model precision. The multilayer perceptron procedure, a class of feed-forward learning model, consists of ≥3 layers of nodes: input, hidden, and output.17 Automatic architecture building, which computes the best number of units in a hidden layer, was performed with SPSS software. Each hidden unit is an activation function of the weighted sum of the inputs; the values of the weights are determined by an estimation algorithm. In this study, the hidden layer consisted of 12 units (nodes). A hyperbolic tangent activation function was also employed for the hidden layers. Softmax activation and cross-entropy error functions were used for the output layer. The multilayer perceptron procedure utilised a backpropagation technique for supervised training. Learning occurred in the recognition phase for each piece of data via changes to connection weights based on the amount of error in the output compared with the expected result (gradient descent method).18

The training process was terminated when no further decreases in calculated error were observed. Subsequently, network weights were identified and used to compute test values. The importance of an independent variable was regarded as a measure of the extent to which network model–predicted values differed from observed values. Normalised importance, expressed as a percentage, constituted the ratio between the importance of each predictor variable and the largest importance value. Model stability was assessed by tenfold cross-validation. Oversampling of minority classes was performed via duplication to manage imbalances in outcome data.

Categorical and continuous variables were expressed as numbers (percentages) and medians (interquartile ranges). The Chi squared test or Fisher’s exact test was used for comparisons of categorical data; the Mann-Whitney U test was used for comparisons of continuous data. The performances of ANN, APACHE II, and APACHE IV models were evaluated in terms of discrimination and calibration power. Discrimination, which constitutes the ability of a predictive model to separate data into classes (eg, death or survival), was evaluated using the area under the receiver operating characteristic curve (AUROC). The AUROCs of the models were compared using the DeLong test. Calibration, which represents the closeness of model probability to the underlying probability of the study population, was evaluated using the Brier score, Hosmer–Lemeshow statistic, and calibration curves.19 All P values were two-sided, and values < 0.05 were considered statistically significant. All analyses were performed with SPSS software and MedCalc statistical software (version 19.6.1).

In total, 14 503 patients were included. The demographic characteristics and hospital mortality data of the study cohort were shown in Table 1, while the physiological and laboratory parameters required to generate an APACHE score were presented in Table 2. Among the recruited patients, 4.93% had at least one missing data point, and the overall rate of missing data was 0.48%. Furthermore, 1400 (9.7%) of the recruited patients were randomly assigned to the validation set; the remaining patients (n=13 103, 90.3%) were assigned to the model development set. With respect to the ANN model, 70% and 30% of the development set were used for training and testing purposes, respectively. The median age was 67 years (interquartile range [IQR]=54-78), median APACHE II score was 18 (IQR=13-25), and median APACHE IV score was 66 (IQR=46-91). The overall hospital and ICU mortality rates were 19.3% (n=2799) and 9.6% (n=1392), respectively.

The baseline co-morbidities, source of admission, disease category, APACHE II score, and APACHE IV score were similar in the test and validation sets (Table 1). More patients in the validation set received continuous renal replacement therapy (18.3% vs 16.1%; P=0.04). Concerning the worst physiological and laboratory parameters within the first 24 hours (Table 2), there were almost no significant differences between the development and validation sets; notably, the haemoglobin level was lower in the validation set (11.3 g/dL vs 11.5 g/dL; P=0.02).

In the development set, the ANN model (AUROC=0.89, 95% confidence interval [CI]=0.88-0.92, Brier score=0.10; P in Hosmer–Lemeshow test=0.34) outperformed the APACHE II model (AUROC=0.80, 95% CI=0.79-0.81, Brier score=0.15; P<0.001) and APACHE IV model (AUROC=0.84, 95% CI=0.83-0.85, Brier score=0.12; P<0.001) for prediction of hospital mortality. The cross-validation accuracy ranged from 0.98 to 1 (mean=0.99), indicating that our ANN model had good stability. There was no statistically significant difference between our ANN model and an ANN model created by oversampling of minority classes (AUROC=0.89, 95% CI=0.89-0.90; P=0.103).

In the validation set, the ANN model (AUROC=0.88, 95% CI=0.86-0.90, Brier score=0.10, P in Hosmer–Lemeshow test=0.37) was superior to the APACHE II model (AUROC=0.85, 95% CI=0.80-0.85, Brier score=0.14; P<0.001 for both comparisons of AUROCs and Brier scores) but similar to the APACHE IV model (AUROC=0.87, 95% CI=0.85-0.89, Brier score=0.11; P=0.34 for comparison of AUROCs, and P=0.05 for comparison of Brier scores) [Fig 1].

The calibration curve for the validation set showed that the ANN model (Fig 2a) outperformed the APACHE IV model (Fig 2b) and the APACHE II model (Fig 2c).

The importances of the predictor variables in predictions of hospital mortality using the ANN model were evaluated. Within 24 hours of ICU admission, the highest sodium level was the most important variable, followed by the highest bilirubin level and the lowest white cell count. Details regarding the normalised importance of each covariate are presented in online supplementary Tables 1 and 2.

To our knowledge, this is the first study in Asia to assess the performance of ANN and compare it with the performances of two extensively validated and benchmark scoring systems—APACHE II and APACHE IV—in terms of predicting hospital mortality among critically ill patients. We found that the ANN model provided better discrimination and calibration compared with the APACHE II model. However, the difference between the ANN and APACHE IV models was less prominent. Calibration was slightly better with the ANN model, but discrimination was similar between the ANN and APACHE IV models.

Conventional logistic regression–based APACHE systems often lose calibration over time and require regular updates to maintain performance.6 11 20 21 22 The original APACHE II model was developed over 30 years ago using data from 13 different hospitals in the US; it was validated in the country before clinical application.2 Studies in Hong Kong7 and Singapore23 have shown that the APACHE II model has good discrimination but poor calibration for ICU patients in Asia. Calibration remained suboptimal regardless of customisation as demonstrated by Lew et al,23 indicating the need for a new prognostic prediction model. Wong and Young24 showed that the APACHE II model had equivalent performance status compared with an ANN model that had been trained and validated using the original APACHE II data. In a medicalneurological ICU in India, an ANN model trained on an Indian population (with or without redundant variables) demonstrated better calibration compared with the APACHE II model.25 The authors speculated that this finding was partly related to differences in standards of care and resources between American and Indian ICUs.25 Overall, differences in case mix, advances in medical technology, and the use of more recent data may explain the superiority of our ANN model compared with the APACHE II model.

Compared with ICU patients in the US, it is fivefold more common for Hong Kong ICU patients to begin renal replacement therapy.26 More than 50% of critically ill patients in Hong Kong require mechanical ventilation, compared with 28% in the US.26 27 A recent population-based study of all patients admitted to adult ICUs in Hong Kong between 2008 and 2018 showed that the APACHE IV standardised mortality ratio decreased from 0.81 to 0.65 during the study period, implying a gradual decline in the performance of the APACHE IV model.26 This model, which was established using data derived from >100 000 ICU patients in 45 US hospitals between 2002 and 2003,5 also tends to overestimate hospital mortality among ICU patients in Hong Kong. In contrast to our study population, where Asian ethnicities were most common, 70% of the patients in APACHE IV reference population were Caucasian.5 The subtle differences in performance between our ANN model and the APACHE IV model could be related to differences in timing during the development of the models. Nevertheless, our ANN model trained on a Hong Kong population was better calibrated for prediction in such a population, compared with the APACHE IV model. This improved calibration could be related to differences in target population (Asian vs Caucasian), epidemiology, and disease profile.

The selection of appropriate variables is a key aspect of model development. The inclusion of additional predictor variables does not necessarily improve a model’s overall performance. Redundant variables may result in overfitting and produce a complicated predictive model without additional benefits. A recently published large national cohort study from Sweden showed that a simplified ANN model with eight parameters outperformed the Simplified Acute Physiology Score III model in terms of discrimination and calibration.28 Among the eight parameters, age and leukocyte count were the most and least important variables, respectively. Notably, leukocyte count was the most important variable in terms of predicting mortality among patients on continuous renal replacement therapy.29 Similar to the present study, Kang et al29 found that age was the 12th most important variable. The overall performance of an ANN model trained with APACHE II parameters in an Indian population could be maintained with the 15 highest information gain variables, including serum sodium level and leukocyte count.25

Among the 53 parameters in our ANN model, the highest sodium level, highest bilirubin level, and lowest white cell count within 24 hours of ICU admission were the top three most important predictor variables (online supplementary Table 1). The association between acquired hypernatraemia and increased hospital mortality among critically patients has consistently been demonstrated in multiple studies.30 31 Hyperbilirubinaemia, another complication in patients with sepsis, was associated with the onset of acute respiratory distress syndrome.32 Sepsis and gastrointestinal/hepatobiliary diseases caused ICU admission in approximately 40% of our patients, possibly explaining the importance of hyperbilirubinaemia in our ANN model. Although the importance of leukocyte count has been demonstrated in other mortality prediction models, the previous models did not specify whether the count was high or low.25 28 29 In the present study, the lowest white cell count was more important than the highest white cell count. Another intriguing observation was that age constituted the 11th most important predictor in our ANN model (online supplementary Table 1). Age is a predictor of survival in many prognostic models.3 5 28 Increasing biological age is often associated with multiple co-morbidities and a progressive decline in physiological reserve, leading to increased mortality. However, a recently published systematic review of 129 studies showed large variations in ICU and hospital mortality rates among older ICU patients, ranging from 1% to 51% in single-centre retrospective studies and 6% to 28% in multicentre retrospective studies.33 These results could be related to differences in admission policies, premorbid functional status, and the intensity of provided to older critically ill patients.

Our ANN model was trained and internally validated on a large number of representative data samples that included most patients admitted to a tertiary ICU in Hong Kong over the past decade. This approach addressed the small sample size limitation that was common in previous studies.24 25 34 All data were automatically collected by a computer system, eliminating the risk of human error during data extraction. Healthcare system digitalisation and advances in information technology have enabled effortless generation of abundant clinical data (eg, physiological parameters, laboratory results, and radiological findings), which can facilitate data collection and development of a new risk prediction model via machine learning.35 36 We hope that generalisability to other ICUs in Asia can be achieved through external validation studies.

This study had some limitations. Although the sample size was large, all data were collected from a single centre; in contrast, data for the APACHE scoring system were derived from multiple large centres. Because the primary objective of the present study was comparison of performance between our ANN model and the APACHE II and APACHE IV models using identical parameters, we did not attempt to determine the optimal subset of parameters that would maintain high ANN performance.25 28 Furthermore, our ANN model may not be applicable to other centres with different case mixes and medical approaches. The lack of external validation may lead to concerns about overfitting, which is a common challenge in ANN model development. Because mortality prediction among ICU patients is a dynamic process, other limitations include the use of static data and the lack of a fixed time point for mortality assessment.

Mortality prediction among critically patients is a challenging endeavour. Our ANN model, which was trained with representative data from a Hong Kong population, outperformed the internationally validated APACHE II model with respect to critically ill patients in Hong Kong. In contrast to the APACHE IV model, our ANN model demonstrated better calibration but similar discrimination performance. External validation studies using data from other hospitals are recommended to confirm our findings. Future studies should explore the feasibility of reducing the number of variables while preserving the discrimination and calibration power of the ANN model. The widespread use of computerised information systems, rather than paper records, in ICU and general ward settings has led to increased data availability. Artificial neural networks, along with other machine learning techniques, are valuable tools that can enhance real-time prognostic prediction.

Concept or design: S Lau, HP Shum, CCY Chan.
Acquisition of data: S Lau, HP Shum, CCY Chan.
Analysis or interpretation of data: S Lau, HP Shum, CCY Chan.
Drafting of the manuscript: S Lau.
Critical revision of the manuscript for important intellectual content: MY Man, KB Tang, KKC Chan, AKH Leung, WW Yan.

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 interests.

Part of the research was presented at the 34th Annual Congress of the European Society of Intensive Care Medicine (3-6 October 2021, virtual) and the Annual Scientific Meeting 2021 of Hong Kong Society of Critical Care Medicine (12 December 2021, virtual).

This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors.

The study protocol complies with the Declaration of Helsinki and was approved by the Hong Kong East Cluster Research Ethics Committee of Hospital Authority, Hong Kong (Ref No.: HKECREC-2021-024). The requirement for patient consent was waived by the Committee due to the retrospective nature of the study.

The supplementary material was provided by the authors and some information may not have been peer reviewed. 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. Keegan MT, Gajic O, Afessa B. Severity of illness scoring systems in the intensive care unit. Crit Care Med 2011;39:163-9. Crossref

2. Breslow MJ, Badawi O. Severity scoring in the critically ill: part 1—interpretation and accuracy of outcome prediction scoring systems. Chest 2012;141:245-52. Crossref

3. Knaus WA, Draper EA, Wagner DP, Zimmerman JE. APACHE II: a severity of disease classification system. Crit Care Med 1985;13:818-29. Crossref

4. APACHE III study design: analytic plan for evaluation of severity and outcome [editorial]. Crit Care Med 1989;17:S169-221.

5. Zimmerman JE, Kramer AA, McNair DS, Malila FM, Shaffer VL. Intensive care unit length of stay: benchmarking based on Acute Physiology and Chronic Health Evaluation (APACHE) IV. Crit Care Med 2006;34:2517-29. Crossref

6. Zimmerman JE, Kramer AA. Outcome prediction in critical care: the Acute Physiology and Chronic Health Evaluation models. Curr Opin Crit Care 2008;14:491-7. Crossref

7. Tan IK. APACHE II and SAPS II are poorly calibrated in a Hong Kong intensive care unit. Ann Acad Med Singap 1998;27:318-22.

8. Gupta R, Arora VK. Performance evaluation of APACHE II score for an Indian patient with respiratory problems. Indian J Med Res 2004;119:273-82.

9. Choi JW, Park YS, Lee YS, et al. The ability of the Acute Physiology and Chronic Health Evaluation (APACHE) IV score to predict mortality in a single tertiary hospital. Korean J Crit Care Med 2017;32:275-83. Crossref

10. Ghorbani M, Ghaem H, Rezaianzadeh A, Shayan Z, Zand F, Nikandish R. A study on the efficacy of APACHEIV for predicting mortality and length of stay in an intensive care unit in Iran. F1000Res 2017;6:2032. Crossref

11. Ko M, Shim M, Lee SM, Kim Y, Yoon S. Performance of APACHE IV in medical intensive care unit patients: comparisons with APACHE II, SAPS 3, and MPM0 III. Acute Crit Care 2018;33:216-21. Crossref

12. Xie J, Su B, Li C, et al. A review of modeling methods for predicting in-hospital mortality of patients in intensive care unit. J Emerg Crit Care Med 2017;1:18. Crossref

13. Tu JV. Advantages and disadvantages of using artificial neural networks versus logistic regression for predicting medical outcomes. J Clin Epidemiol 1996;49:1225-31. Crossref

14. Sargent DJ. Comparison of artificial neural networks with other statistical approaches: results from medical data sets. Cancer 2001;91(8 Suppl):1636-42. Crossref

15. Meiring C, Dixit A, Harris S, et al. Optimal intensive care outcome prediction over time using machine learning. PLoS One 2018;13:e0206862. Crossref

16. Rajkomar A, Oren E, Chen K, et al. Scalable and accurate deep learning with electronic health records. NPJ Digit Med 2018;1:18. Crossref

17. Norgaard M, Ravn O, Poulsen NK, Hansen LK. Neural Networks for Modelling and Control of Dynamic Systems: A Practitioner’s Handbook. London: Springer; 2000.

18. Ludermir TB, Yamazaki A, Zanchettin C. An optimization methodology for neural network weights and architectures. IEEE Trans Neural Netw 2006;17:1452-9. Crossref

19. Steyerberg EW, Vickers AJ, Cook NR, et al. Assessing the performance of prediction models: a framework for traditional and novel measures. Epidemiology 2010;21:128-38. Crossref

20. Lam KW, Lai KY. Evaluation of outcome and performance of an intensive care unit in Hong Kong by APACHE IV model: 2007-2014. J Emerg Crit Care Med 2017;1:16. Crossref

21. Paul E, Bailey M, Van Lint A, Pilcher V. Performance of APACHE III over time in Australia and New Zealand: a retrospective cohort study. Anaesth Intensive Care 2012;40:980-94. Crossref

22. Mann SL, Marshall MR, Holt A, Woodford B, Williams AB. Illness severity scoring for intensive care at Middlemore Hospital, New Zealand: past and future. N Z Med J 2010;123:47-65.

23. Lew CC, Wong GJ, Tan CK, Miller M. Performance of the Acute Physiology and Chronic Health Evaluation II (APACHE II) in the prediction of hospital mortality in a mixed ICU in Singapore. Proc Singapore Healthc 2019;28:147-52. Crossref

24. Wong LS, Young JD. A comparison of ICU mortality prediction using the APACHE II scoring system and artificial neural networks. Anaesthesia 1999;54:1048-54. Crossref

25. Nimgaonkar A, Karnad DR, Sudarshan S, Ohno-Machado L, Kohane I. Prediction of mortality in an Indian intensive care unit. Comparison between APACHE II and artificial neural networks. Intensive Care Med 2004;30:248-53. Crossref

26. Ling L, Ho CM, Ng PY, et al. Characteristics and outcomes of patients admitted to adult intensive care units in Hong Kong: a population retrospective cohort study from 2008 to 2018. J Intensive Care 2021;9:2. Crossref

27. Wunsch H, Angus DC, Harrison DA, Linde-Zwirble WT, Rowan KM. Comparison of medical admissions to intensive care units in the United States and United Kingdom. Am J Respir Crit Care Med 2011;183:1666-73. Crossref

28. Holmgren G, Andersson P, Jakobsson A, Frigyesi A. Artificial neural networks improve and simplify intensive care mortality prognostication: a national cohort study of 217,289 first-time intensive care unit admissions. J Intensive Care 2019;7:44. Crossref

29. Kang MW, Kim J, Kim DK, et al. Machine learning algorithm to predict mortality in patients undergoing continuous renal replacement therapy. Crit Care 2020;24:42. Crossref

30. O’Donoghue SD, Dulhunty JM, Bandeshe HK, Senthuran S, Gowardman JR. Acquired hypernatraemia is an independent predictor of mortality in critically ill patients. Anaesthesia 2009;64:514-20. Crossref

31. Olsen MH, Møller M, Romano S, et al. Association between ICU-acquired hypernatremia and in-hospital mortality: data from the medical information mart for intensive care III and the electronic ICU collaborative research database. Crit Care Explor 2020;2:e0304. Crossref

32. Zhai R, Sheu CC, Su L, et al. Serum bilirubin levels on ICU admission are associated with ARDS development and mortality in sepsis. Thorax 2009;64:784-90. Crossref

33. Vallet H, Schwarz GL, Flaatten H, de Lange DW, Guidet B, Dechartres A. Mortality of older patients admitted to an ICU: a systematic review. Crit Care Med 2021;49:324-34. Crossref

34. Clermont G, Angus DC, DiRusso SM, Griffin M, Linde-Zwirble WT. Predicting hospital mortality for patients in the intensive care unit: a comparison of artificial neural networks with logistic regression models. Crit Care Med 2001;29:291-6. Crossref

35. Kim S, Kim W, Park RW. A comparison of intensive care unit mortality prediction models through the use of data mining techniques. Healthc Inform Res 2011;17:232-43. Crossref

36. Bulgarelli L, Deliberato RO, Johnson AE. Prediction on critically ill patients: the role of “big data”. J Crit Care 2020;60:64-8. Crossref

The coronavirus disease 2019 (COVID-19) pandemic has affected >770 million people worldwide, causing >7 million deaths as of 31 December 2023.1 The period between July and September 2020 constituted the third wave of the pandemic in Hong Kong, resulted in >1.2 million reported cases between 23 January 2020 and 29 January 2023.2 The containment strategies implemented during the third wave included mandatory mask wear in public places, even when exercising in public outdoor areas; suspensions of public leisure facilities and private gyms; and the initiation of work-from-home arrangements.3 These strategies led to reductions in physical activity and daily movement, with the goal of viral containment. Furthermore, compulsory social distancing and suspension of dine-in services were included among the policies that could affect various dietary and lifestyle habits, although these methods were less stringent than approaches in cities under lockdown. Overall, the unprecedented public health crisis created many challenges for Hong Kong residents attempting to maintain healthy lifestyle habits. Nevertheless, few studies have examined dietary and physical activity habits in the general population during the COVID-19 pandemic.4 5 6

Considering that individuals with chronic diseases are more likely to develop severe cases of COVID-19, this study aimed to measure the prevalences of unhealthy dietary habits and physical inactivity levels in an adult Chinese population, to identify factors associated with their adoption of these dietary and physical activity habits, and to perform a time trend analysis comparing the proportions of the population that adopted healthy dietary habits, physical activity levels, and avoidance of smoking and alcohol consumption during the third wave of the COVID-19 pandemic.

We utilised a methodology similar to a previous population-based, random telephone survey conducted in Hong Kong.7 Two-stage sampling was performed, in which participants were recruited by trained interviewers through a telephone interview system based on telephone calls to landlines identified by random digit dialling. The sample population was randomly selected by the Centre for Health Behaviours Research at The Chinese University of Hong Kong. Calls were made during typical office hours, 9 am to 5 pm, Monday through Saturday between 7 and 31 October 2020. Three attempts were made if the call initially was not answered. Territory-wide, any Chinese adults aged ≥18 years who could communicate in Chinese via telephone were eligible to participate. Assuming an outcome variable rate of 35%, at least 1456 participants were required to achieve a precision level of 2.5% from the following formula:
where ‘p’ stands for proportion and ‘N’ stands for sample size.

The interviews were performed using a fieldwork manual highlighting standard operating procedures by a team of trained interviewers and supervised by an experienced project coordinator throughout the study. The characteristics of survey participants are shown in Table 1.

The survey consisted of five sections: (1) socio-demographic details (age, sex, marital status, education level, job status, household income, and receipt of comprehensive social security assistance); (2) clinical information (eg, presence of chronic diseases); (3) smoking (current daily amount/ex-/non-smoker) and alcohol consumption habits (daily amount in the preceding 7 days); (4) dietary screening via the Hong Kong Diet Score (HKDS), using a validated scale that contained nine items assessing the participant’s daily consumption of nine food groups in the preceding 7 days; and (5) level of physical activity in the preceding 7 days, as determined by a Chinese version of the 7-item International Physical Activity Questionnaire Short Form (IPAQ-C).

The traditional Mediterranean diet is well-defined and has been positively associated with favourable health outcomes.8 9 10 The Mediterranean diet score is used to measure compliance with a traditional Mediterranean diet. This scoring system has been widely utilised in studies that measure Mediterranean diet adherence or adaptation as an indicator of healthy dietary choices. In this study, we developed the HKDS, a dietary screener that contained nine items assessing dietary intake of nine food groups (alcohol, legumes, grains, fruits, vegetables, meats, dairy, red/orange vegetables, and fatty fish) in the preceding 7 days. The screener incorporated key traditional Greek diet characteristics, known as the Mediterranean diet score of de Groot et al,8 which were also used in a study of Hong Kong Chinese by Woo et al (Table 2).11 The original 8-item survey was modified by removing the ratio of monounsaturated fatty acids to saturated fatty acids and replacing ethanol with alcohol. Dietary fatty acids and ethanol are widely distributed among various food groups; they are typically assessed through weighted foods, which are unlikely to be accurately determined using a single question in a telephone interview. Two additional items were included regarding carotenoid-rich and omega-3–rich food intake based on the Hong Kong Centre for Food Safety Recommended Nutrient Intake for vitamin A12 and the World Health Organization recommendation for omega-3. Both nutrients are inversely associated with incidence of non-communicable diseases (NCDs). For each item, consumption at or above the recommended amount was scored as 1 point and 0 points otherwise; however, for ethanol, consumption below the specified amount was scored as 1 point and 0 points otherwise. Each participant received a total score of 0 to 9; a score of ≥5 was considered high. A pilot survey was conducted with a convenience sample of 23 participants. Intraclass correlation coefficient estimates and 95% confidence intervals (CIs) were determined using a two-way mixed-effects model to assess internal consistency regarding the number of serves (ie, serving sizes of the food group consumed) reported. The intraclass correlation coefficient was 0.87, indicating good reliability. Cohen’s κ was calculated to evaluate agreement between test and retest scores. Agreement between the two tests was fair (κ=0.24, 95% CI=-0.15 to 0.63; P=0.239).

The IPAQ-C score was regarded as a categorical variable indicating exercise level based on the frequency and intensity of physical activity: (1) low (total activity <600 metabolic equivalent of task [MET]–minutes/week), (2) moderate (total activity ≥600 MET-minutes/week), or (3) high (total activity >3000 MET-minutes/week).

Finally, an unhealthy lifestyle score (0 to 4) was assigned to each participant based on a composite outcome involving low HKDS, physical inactivity, current smoking habit, and alcohol consumption; each unhealthy habit contributed 1 point to the score.

We used SPSS software (Windows version 26.0; IBM Corp, Armonk [NY], United States) for data analysis. Descriptive analyses were performed regarding the participants’ socio-demographic details, clinical information (eg, presence of chronic diseases), and the HKDS. The primary outcome variables included: (1) unhealthy dietary habits (low HKDS score); (2) suboptimal physical activity (low IPAQ-C score, indicating low exercise level); and (3) unhealthy lifestyle score (≥2). Univariable logistic regression was performed to examine associations between socio-demographic variables and each of the first two outcome variables. Multivariable logistic regression was modelled by controlling for covariates with P values <0.20 in univariable regression analysis, a cut-off level commonly used in public health research. For example, Torenfält and Dimberg13 utilised this approach when evaluating stroke and death in middle-aged Swedish men. The approach was also used in a French study14 concerning medical features of patients with COVID-19 and influenza. Additionally, linear regression analysis was conducted in the present study to examine associations between socio-demographic variables and the unhealthy lifestyle score. Time trends for various food intake, physical inactivity, current smoking, and alcohol consumption statuses were evaluated; the prevalences of these lifestyle habits were compared with population-wide figures from governmental reports over the past two decades using the Chi squared test for heterogeneity. P values <0.05 were considered statistically significant.

The Centre for Health Protection has been conducting health surveys periodically since 2003 to collect information about health and lifestyle-related behaviours, as well as practices related to the prevention of NCDs among residents aged ≥15 years.15 The resulting reports have presented key findings concerning physical activity, dietary habits, alcohol consumption, and smoking habits, as well as other self-care practices. We gathered relevant findings from 14 governmental reports covering the period from 2003-2004 to 2018-2019 (calendar years with the most updated figures)15 to perform trend analysis of fruit and vegetable consumption, physical activity level, smoking, and alcohol consumption status among Hong Kong residents. These results were compared with the findings of the present study; adjustments were solely performed for sex because the age distribution was limited in all but the most recent reports.

In total, 2551 individuals were contacted for a telephone interview and 1500 participated; the response rate was 58.8%. Most interviewed individuals were older adults (≥65 years, 66.7%), women (65.6%), and married (77.9%). Of the participants, 40.1% were engaged in professional and office work; only 16.3% had attained a tertiary education or higher. About 16% of participants reported a household income ≥HK$30 000, whereas 41.1% reported a household income <HK$10 000. Health status was predominately self-reported as average (42.5%) or above average (51.6%). More than half of the participants (54.5%) were undergoing long-term medical management or were taking medications for chronic diseases; the most common chronic conditions were diabetes mellitus (21.0%) and hypertension (42.7%) [Table 1].

Dietary habits, as measured by the HKDS (score range, 2-9), were classified as high scoring (5-9) or low scoring (0-4). Approximately 51% of participants had a low score (Table 1), suggesting minimal adherence to the traditional Chinese eating pattern; these participants were mainly younger individuals (aged ≤34 years, 66%) and men (58%) [Table 2]. Greater proportions of participants with lower income, current smokers, and current drinkers had low scores according to the HKDS (54%, 69%, and 67%, respectively) [Table 2]. Participants with chronic diseases had various HKDS results; <50% of patients with renal diseases and diabetes had a high score, and this result indicated that they had a poor dietary habits.

A greater risk of practising unhealthy dietary habits (low HKDS) was associated with male sex (adjusted odds ratio [aOR]=1.31, 95% CI=1.03-1.67), non-married status (ie, single/divorced/widowed) [aOR=1.56, 95% CI=1.20-2.03], a diagnosis of diabetes (aOR=1.53, 95% CI=1.15-2.03), and alcohol consumption (aOR=1.76, 95% CI=1.17-2.64) [Table 3].

Among all participants, 35.5%, 54.0% and 10.5% had low, moderate, and high levels of physical activity, respectively (Table 1). Participants who were non-married (aOR=1.66, 95% CI=1.23-2.22), undergoing long-term management of medical diseases (aOR=1.65, 95% CI=1.08-2.54), had diabetes (aOR=1.39, 95% CI=1.02 to 1.89), and had renal diseases (aOR=9.32, 95% CI=2.06-42.25) exhibited greater likelihood of physical inactivity (Table 4).

Few participants were current daily smokers (2.8%) and alcohol drinkers (8.5%) [Table 1].

Factors associated with an unhealthy lifestyle score are presented in Table 5. Male sex (beta coefficient [β]=0.25, 95% CI=0.16-0.34), non-married status (β=0.19, 95% CI=0.08-0.29), manual work (β=0.17, 95% CI=0.07-0.27), self-reported poor or very poor health status (β=0.26, 95% CI=0.07-0.45), a diagnosis of diabetes (β=0.31, 95% CI=0.20-0.41), and a diagnosis of renal disease (β=0.92, 95% CI=0.53-1.31) increased the likelihood of poor lifestyle habits. Housewife or retired status (β=-0.15, 95% CI=-0.25 to -0.04) and a higher household income (≥HK$30 000; β=-0.20, 95% CI=-0.33 to -0.08) decreased the likelihood of poor lifestyle habits.

The Health Behaviour Survey, with a response rate of 70.8% in the 2018/2019 report, is a population-based fieldwork study conducted by the Centre for Health Protection of the Department of Health.16 In that survey, female participants comprised 52.7% of the sample, compared with 65.6% in the present telephone survey. The age-group with the largest proportion of participants in the Survey was 65 to 74 years (36.7% vs 11.0%), which might have influenced the sex ratio.

Time trend analysis showed that the proportion of surveyed Hong Kong residents eating five daily servings of fruits and vegetables declined for both sexes in general (Fig a). Similarly, a significantly smaller proportion of participants reported walking >10 minutes for ≥5 days per week, and this proportion has continued to decline since 2016 (Fig b). There was a gradual decrease in the number of participants with a moderate or high level of physical activity. Despite a notable peak in 2019, there was a decline in 2020 with <60% of participants reportedly engaging in these physical activities. Finally, significantly smaller proportion of the study participants reported not currently smoking or consuming alcohol, compared with the proportions in previous population-based surveys (2010-2019) [Fig c and d].

In this population-based study of 1500 Hong Kong residents during the third wave of the COVID-19 pandemic, we found that the proportion of people with healthy food intake (ie, daily consumption of five servings of fruits and vegetables) has decreased since 2003; although a slight increased was observed in 2020, it was still below the overall average. Additionally, we found that the prevalence of low physical activity has gradually increased. In contrast, the rates of smoking and alcohol consumption were below the rates observed in pre-pandemic population-based surveys. Men and women in various age-groups had dietary habits less adherent to the traditional Chinese eating pattern, as measured by the HKDS, than 20 years prior. Adherence to the traditional eating pattern was significantly lower among male participants, single individuals, low-income participants, alcohol drinkers, individuals with low physical activity, and patients with diabetes mellitus or renal disease. We also found that men, individuals with non-married status, manual workers, individuals with self-perceived poor or very poor health status, and patients with diabetes and renal disease had a greater likelihood of poor lifestyle habits.

Worldwide, insufficient intake of fruits and vegetables and inadequate physical activity have been attributed to 34% and ≥20% of NCDs, respectively.17 18 The incidence of chronic diseases (eg, cancers, diabetes, and cardiovascular disease) in Hong Kong has increased by 60% in the past two decades.19 Although processed food intake was not assessed in this study, the HKDS results indicated very low intakes of legumes and red/orange vegetables among Hong Kong residents; a previous study showed that fruit intake in Hong Kong is among the lowest levels worldwide.20 The report of Health Behaviour Survey 2018/2019 revealed that 95.6% of surveyed participants had inadequate daily fruit and vegetable intake, based on World Health Organization recommendations.16 The present study showed further reduction, such that the proportion of Hong Kong residents consuming the recommended (≥5) daily servings of fruits and vegetables declined by 3.9% in the past two decades. Additionally, 10% of the population reported consuming more than one daily serving of processed meat, and the frequency of processed meat consumption increased by 3.1%.16 Another known risk factor for NCDs, overweight or obesity, affected approximately half of the Hong Kong population aged 15 to 84 years in 2015; this value was slightly (2%) below the global average.21

Disruptions to usual routines can lead to new dietary behaviours. For example, dine-in restrictions at restaurants and bars during the pandemic led to greater use of food delivery services, and take-away food is often less healthy.22 However, we observed increased fruit and vegetable consumption after the third wave of the COVID-19 outbreak. This finding is supported by the work of Wang et al,23 which showed that fruit and vegetable consumption increased. Collectively, the disease prevention and control policies that prohibited group gatherings and shortened the operating hours of bars and clubs (if such facilities were not entirely closed) could have significantly reduced smoking and alcohol consumption in Hong Kong during the pandemic. Additionally, the lower level of physical activity could be related to the closure of public sports avenues and restricted access to sports facilities.

The dietary habits of Hong Kong residents have changed from the traditional Chinese diet to a fast-paced dining experience involving convenient, processed foods with limited diversity. Hong Kong has a population of >7.5 million people, and 90% of its food supply is imported from other countries.24 Woo et al11 concluded that, despite geographical and cultural differences, traditional Chinese dietary habits were conceptually similar to the health-promoting Mediterranean diet. In 2001, high overall adherence to the Mediterranean diet was observed across all age-groups in Hong Kong, except for younger populations and men.11 In contrast, the present study showed that a smaller proportion of participants in all age and sex groups had high overall adherence to the Mediterranean diet. In this study, intriguingly, women consistently exhibited higher Mediterranean diet score (71%) and HKDS (53%; P<0.005) results, compared with their male counterparts. Moreover, the present study revealed lower prevalences of some chronic diseases among women: diabetes, cardiovascular disease, liver disease, and renal disease. The contemporary diets in modern Hong Kong and many developed regions have low fibre content and high processed food content; they also include food additives, refined sugar, and hydrogenated fats.25 The subtle but consistent westernisation of dietary habits appears to be detrimental for residents who consistently consume these food items.

A lack of colourful vegetables and fruits may reduce the diversity of beneficial gut microbiota,26 although leafy greens such as pak choy (Chinese cabbage), choy sum (Chinese flowering cabbage), and Chinese kale are staple foods in the Hong Kong diet throughout the year. Moreover, in a review of literature concerning exercise and gut microbial composition, Mitchell et al27 found that exercise alters gut microbiota; however, the direction of apparent change has varied among studies. Increases in butyrate-producing bacteria and faecal butyrate concentrations, with protective anti-inflammatory effects and the potential to enhance anti-infection immunity, have been observed among physically active adults.28 The mechanisms are unclear but the benefits of adopting a lifestyle with a diverse diet and physical activity consistently create an optimal environment for gut microbiota.

The large sample size and random sampling design of this territory-wide survey were strengths that enhanced the validity of the findings. However, this study had several limitations that should be addressed. First, cause-and-effect relationships between the COVID-19 pandemic and changes in lifestyle habits could not be established because of the cross-sectional approach. Individuals in home quarantine during the third wave of the pandemic might have experienced temporary changes in lifestyle habits; a prospective observational study and longer trend analysis are needed to facilitate long-term comparisons. Factors other than the pandemic (eg, mental wellness) could also have affected lifestyles among the study population. Although the present study utilised a random sampling strategy, non-response and selection biases were possible because younger segments of the Hong Kong population did not use landlines during the study period. Furthermore, response and social desirability biases may have been present in this telephone survey. Nevertheless, the high response rate and anonymous nature of this survey may have minimised these potential biases. Additionally, generalisation of the study findings should be performed with caution because the survey only included a single Chinese population. Considering that the participants’ characteristics differed from the general population, the findings might not be directly applicable to the general public. Moreover, the survey was conducted in 2020, and lifestyle habits in the general population might have changed throughout the pandemic. Finally, the low case numbers of some self-reported diseases, such as renal disease and cancer, may have resulted in type II error.

This representative population-based survey revealed that larger proportions of the general population had unhealthy lifestyles, including dietary habits and physical inactivity, during the COVID-19 pandemic than during pre-pandemic years. There is an urgent need to formulate and implement effective public health strategies at both individual and organisational levels. The encouragement of healthy lifestyles through evidence-based health promotion programmes is essential. This encouragement could be conveyed to communities through organised and concerted efforts by the government and relevant stakeholders. Future studies should evaluate the effectiveness of various interventions and approaches to achieve these important goals.

Concept or design: WYY Lin, MCS Wong.
Acquisition of data: WYY Lin, J Huang, Y Bai.
Analysis or interpretation of data: WYY Lin, J Huang, Y Bai.
Drafting of the manuscript: WYY Lin, MCS Wong.
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 editors of the journal, MCS Wong and J Huang were not involved in the peer review process. Other authors have disclosed no conflicts of interest.

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 Survey and Behavioural Research Ethics Committee of The Chinese University of Hong Kong (Ref No.: SBRE-20-099). The interviewees provided informed consent after they were briefed on the study purpose and being assured of the confidentiality measures in place.

1. World Health Organization. WHO Coronavirus Disease (COVID-19) Dashboard. Available from: https://covid19.who.int/. Accessed 19 Jan 2024.

2. Centre for Health Protection, Department of Health, Hong Kong SAR Government. Situation of COVID-19 (23 January 2020 to 29 January 2023). 2023 July 27. Available from: https://www.chp.gov.hk/files/pdf/local_situation_covid19_en.pdf. Accessed 2 Feb 2024.

3. Chief Executive’s Office, Hong Kong SAR Government. Fighting the virus for six months battling another wave of the epidemic together. 2020 Jul 25. Available from: https://www.ceo.gov.hk/archive/5-term/eng/pdf/article20200725.pdf. Accessed 6 Apr 2022.

4. Mattioli AV, Sciomer S, Cocchi C, Maffei S, Gallina S. Quarantine during COVID-19 outbreak: changes in diet and physical activity increase the risk of cardiovascular disease. Nutr Metab Cardiovasc Dis 2020;30:1409-17. Crossref

5. Bin Mahfoz TM, Shaik RA, Ahmad MS, Alzerwi NA, Almutairi R. A nationwide survey to assess COVID-19’s impact on health and lifestyle in Saudi Arabia. Eur Rev Med Pharmacol Sci 2022;26:4092-101. Crossref

6. Mitchell JJ, Bu F, Fancourt D, Steptoe A, Bone JK. Longitudinal associations between physical activity and other health behaviours during the COVID-19 pandemic: a fixed effects analysis. Sci Rep 2022;12:15956. Crossref

7. Wang JX, Huang J, Cheung CS, Wong WN, Cheung NT, Wong MC. Adoption of an electronic patient record sharing pilot project: cross-sectional survey. J Med Internet Res 2020;22:e13761. Crossref

8. de Groot LC, van Staveren WA, Burema J. Survival beyond age 70 in relation to diet. Nutr Rev 1996;54:211-2. Crossref

9. Fung TT, Rexrode KM, Mantzoros CS, Manson JE, Willett WC, Hu FB. Mediterranean diet and incidence of and mortality from coronary heart disease and stroke in women. Circulation 2009;119:1093-100. Crossref

10. Bamia C, Lagiou P, Buckland G, et al. Mediterranean diet and colorectal cancer risk: results from a European cohort. Eur J Epidemiol 2013;28:317-28. Crossref

11. Woo J, Woo KS, Leung SS, et al. The Mediterranean score of dietary habits in Chinese populations in four different geographical areas. Eur J Clin Nutr 2001;55:215-20. Crossref

12. Centre for Food Safety, Hong Kong SAR Government. Nutrient and health—maintain optimal nutrient intake. 2008. Available from: https://www.cfs.gov.hk/english/multimedia/multimedia_pub/multimedia_pub_fsf_28_02.html. Accessed 30 Sep 2020. Crossref

13. Torenfält I, Dimberg L. Stroke and death — findings from a 25-year follow-up of a cohort of employed Swedish middle-aged men of the Coeur study. J Public Health 2022;30:1713-24. Crossref

14. Faury H, Courboulès C, Payen M, et al. Medical features of COVID-19 and influenza infection: a comparative study in Paris, France. J Infect 2021;82:e36-9. Crossref

15. Center for Health Protection, Department of Health, Hong Kong SAR Government. Non-communicable Disease Branch. Publications. Available from: https://www.chp.gov.hk/en/resources/29/100057.html. Accessed 31 Jan 2024.

16. Centre for Health Protection, Department of Health, Hong Kong SAR Government. Report of Health Behaviour Survey 2018/19. Available from: https://www.chp.gov.hk/files/pdf/report_of_health_behaviour_survey_2018_en.pdf. Accessed 29 Jan 2021.

17. World Health Organization. Global health risks: mortality and burden of disease attributable to selected major risks. 2009. Available from: https://apps.who.int/iris/handle/10665/44203. Accessed 29 Jan 2021.

18. World Health Organization. Physical activity. 2022. Available from: https://www.who.int/news-room/fact-sheets/detail/physical-activity. Accessed 22 Jan 2024.

19. Hospital Authority. Overview of Hong Kong Cancer Statistics of 2018. 2020. Available from: https://www3.ha.org.hk/cancereg/pdf/overview/Overview%20of%20HK%20Cancer%20Stat%202018.pdf. Accessed 30 Sep 2020.

20. Micha R, Khatibzadeh S, Shi P, et al. Global, regional and national consumption of major food groups in 1990 and 2010: a systematic analysis including 266 country-specific nutrition surveys worldwide. BMJ Open 2015;5:e008705. Crossref

21. World Health Organization. Obesity and overweight. 2021. Available from: https://www.who.int/news-room/fact-sheets/detail/obesity-and-overweight. Accessed 23 Jun 2021.

22. Nguyen TH, Vu DC. Food delivery service during social distancing: proactively preventing or potentially spreading coronavirus disease-2019? Disaster Med Public Health Prep 2020;14:e9-10. Crossref

23. Wang J, Yeoh EK, Yung TK, et al. Change in eating habits and physical activities before and during the COVID-19 pandemic in Hong Kong: a cross-sectional study via random telephone survey. J Int Soc Sports Nutr 2021;18:33. Crossref

24. Research Office, Legislative Council Secretariat. Fact sheet: regulation of imported food in Hong Kong. 2016. Available from: https://www.legco.gov.hk/research-publications/english/1516fsc14-regulation-of-imported-food-in-hong-kong-20160226-e.pdf. Accessed 22 Jan 2024.

25. Jew S, AbuMweis SS, Jones PJ. Evolution of the human diet: linking our ancestral diet to modern functional foods as a means of chronic disease prevention. J Med Food 2009;12:925-34. Crossref

26. Garcia-Mantrana I, Selma-Royo M, Alcantara C, Collado MC. Shifts on gut microbiota associated to Mediterranean diet adherence and specific dietary intakes on general adult population. Front Microbiol 2018;9:890. Crossref

27. Mitchell CM, Davy BM, Hulver MW, Neilson AP, Bennett BJ, Davy KP. Does exercise alter gut microbial composition? a systematic review. Med Sci Sports Exerc 2019;51:160-7. Crossref

28. Allen JM, Mailing LJ, Niemiro GM, et al. Exercise alters gut microbiota composition and function in lean and obese humans. Med Sci Sports Exerc 2018;50:747-57. Crossref

Vaccination is an effective tool to combat the coronavirus disease 2019 (COVID-19) pandemic. Two types of COVID-19 vaccines are used in Hong Kong: the Sinovac-CoronaVac inactivated severe acute respiratory syndrome coronavirus 2 vaccine (Sinovac Biotech Ltd, Beijing, China) and Pfizer BioNTech BNT162b2 (Pfizer Inc, Philadelphia [PA], United States) messenger RNA vaccine began distribution on 26 February 2021 and 10 March 2021, respectively.

According to the World Health Organization, vaccine hesitancy is defined as delaying or refusing vaccination despite the availability of vaccination services.1 In a study conducted during the third wave of COVID-19 in Hong Kong, the overall vaccine acceptance rate was approximately 37%.2 Although the subsequent acceptance rate has varied with pandemic progression, confidence in COVID-19 vaccines has remained a key factor in reducing vaccine hesitancy.3

Pregnant women were generally excluded from clinical trials focusing on the development, safety, and efficacy of COVID-19 vaccines.4 When COVID-19 vaccines were introduced in Hong Kong, routine vaccination was not recommended for women who were pregnant or breastfeeding, except when there was a high risk of exposure or complications.5 The relative lack of data may have contributed to vaccine hesitancy among pregnant women.6 7 8 Based on data concerning the efficacy and safety of COVID-19 vaccination in preventing serious illness,9 COVID-19 vaccination is recommended for people who are pregnant, breastfeeding, planning to become pregnant, or may become pregnant in the future.10 11

On 23 April 2021, the Hong Kong College of Obstetricians and Gynaecologists (HKCOG) issued an interim recommendation that pregnant women receive the BioNTech COVID-19 vaccine at the same time as the general population.12 On 18 February 2022, the Sinovac vaccine was also recommended for use in pregnant women.12 Furthermore, the recommended interval between the second and third doses of COVID-19 vaccine was shortened from 180 days to 90 days, beginning on 4 March 2022. The vaccine pass policy for entry to specific premises was tightened on 31 May 2022.13 For persons who were over 18 years old and had no history of infection, a minimal of two doses of vaccination was required. A third dose was required if the second dose was taken over 6 months ago. Starting from 13 June 2022, women attending obstetric clinics were required to provide a negative result proof of a polymerase chain reaction–based nucleic acid test conducted with specimen collected within 48 hours before the visit if they did not fulfil the vaccine pass requirement.14

The fifth wave of COVID-19 in Hong Kong has resulted in an overwhelming number of COVID-19 cases. Pregnant women are not automatically protected from COVID-19; indeed, their vaccine hesitancy and low vaccination rate may lead to greater susceptibility. Information concerning the patterns of COVID-19 vaccination and disease transmission among pregnant women in Hong Kong is unavailable. This study examined patterns of vaccination and transmission among pregnant women who were booked for delivery at a tertiary hospital in Hong Kong, with the goal of providing insights into maternal disease characteristics.

This retrospective review included women who were booked for delivery at Queen Mary Hospital in Hong Kong and had attended the booking antenatal visit from 1 July 2021 to 31 March 2022. Information concerning COVID-19 vaccination history was retrieved from the Clinical Management System of the Hospital Authority, which captured COVID-19 vaccination data from the Department of Health.

Pregnant women were diagnosed with COVID-19 because of symptoms or (in the absence of symptoms) during admission screening. Women diagnosed with COVID-19 through other channels were able to reschedule their appointments. Phone consultations were provided by the obstetric team at Queen Mary Hospital. The clinical details of COVID-19 cases were documented in the computerised Antenatal Record System. Additionally, antenatal progress notes were updated if a pregnant woman reported a history of COVID-19 during a follow-up visit. Data regarding demographic characteristics, COVID-19 status, and obstetric outcomes were retrieved from the Antenatal Record System.

The vaccinated group comprised women who received at least one dose of any type of COVID-19 vaccine. Vaccination periods were classified as pre-pregnancy, antenatal, and postnatal for women with a known date of delivery, miscarriage, or termination of pregnancy as of 30 June 2022. For women with ongoing pregnancies and unknown obstetric outcomes, the vaccination period was estimated according to the expected date of delivery. Antenatal status was regarded as known ongoing pregnancy before 42 weeks of gestation. A vaccination episode was defined as any episode of COVID-19 vaccination including the first, second, and third doses. The number of days elapsed since vaccination was defined as the interval between the last dose of COVID-19 vaccine and 30 June 2022 for women who had received one or two doses of vaccine. Complete vaccination was regarded as the period between 15 and 90 days after the second dose of COVID-19 vaccine, or 14 days after the third dose of COVID-19 vaccine for women who had never been diagnosed with COVID-19. For women with COVID-19, the date of diagnosis was regarded as the reference point when determining vaccination status.

Descriptive statistics were reported. Vaccination rates were calculated according to age-group. Background demographic characteristics were compared between vaccinated and unvaccinated groups. Student’s t test, analysis of variance, and the Chi squared test were used as appropriate. Regression analyses were conducted to identify factors affecting vaccine acceptance. P values <0.05 were considered statistically significant. Statistical analyses were performed using SPSS software (Windows version 26; IBM Corp, Armonk [NY], United States).

Table 1 shows the demographic characteristics of 2396 pregnant women who had attended the booking antenatal visit between 1 July 2021 and 31 March 2022. As of 30 June 2022, 2006 (83.7%), 1843 (76.9%), and 831 (34.7%) women had received the first, second, and third doses of COVID-19 vaccine, respectively. Among the 1843 women who had received two doses of vaccine, 1056 (57.3%) underwent vaccination before pregnancy (Fig 1). Of these 1843 women, 831 received a third dose; the median interval between the second and third doses was 280 days (interquartile range, 239-308). Of the remaining 1012 women who had received only two doses of vaccine, 684 (67.6%) and 504 (49.8%) had already passed the 90-day and 180-day intervals, respectively. Their median number of days elapsed since the last vaccine was 315 (interquartile range, 145-368), which considerably exceeded the recommended 90-day interval.

Only 26.6% (1243/4680) of vaccination episodes occurred during pregnancy. Among women who underwent antenatal vaccination, 65.7% (817/1243) had it during the fifth wave of COVID-19 between January 2022 and June 2022; 65.7% (537/817) of these women received the third dose. The two peaks of vaccination for third dose were observed in early March 2022 and late May 2022 (Fig 2).

The vaccination rate was the lowest among Chinese women (81.4%), but the highest among Caucasian women (96.4%) [Fig 3]. Multivariate analysis showed that active working status (odds ratio [OR]=1.94; 95% confidence interval [CI]=1.47-2.56) was significantly associated with a higher COVID-19 vaccination rate, whereas Chinese ethnicity (OR=0.21; 95% CI=0.13-0.33) and women with obstetric complications (OR=0.72; 95% CI=0.55-0.94) were significantly associated with a lower COVID-19 vaccination rate.

In total, there were 265 (11.1%) COVID-19 cases in this cohort; the earliest diagnosis was made on 1 January 2022 during the fifth wave of COVID-19 in Hong Kong (Table 2). The disease rate was more than tenfold higher in women who had no (20.3%) or incomplete (18.8%) vaccination, compared with women who had complete vaccination (2.1%; P<0.001). After exclusion of pregnancies among women aged ≤19 years, there was an insignificant trend of lower vaccination among young women (ie, aged 20-29 years), but their disease rate was the highest (19.1%) [Fig 4].

Among 237 (89.4%) women who had an antenatal diagnosis of COVID-19, 42 (17.7%) required admission for monitoring and 26 (11.0%) delivered in an isolation facility. No women required intensive care or oxygen support. There were no adverse maternal outcomes or cases of vertical transmission.

To our knowledge, this is the first report of the low COVID-19 vaccination rate (83.7%) among pregnant women in Hong Kong. This rate is considerably lower than the single-dose rate among the general public (92.7%) that was reported on the government’s vaccination dashboard on the final day of the study period (ie, 30 June 2022).15 Furthermore, the disease rate was more than threefold higher in women who had no or incomplete vaccination, compared with women who had complete vaccination.

Pregnant women are considered a vulnerable group. The substantial increase in the disease rate, combined with a lower vaccination rate, among women aged 20 to 29 years is particularly concerning. A case of COVID-19 during pregnancy can lead to adverse obstetric outcomes, including increased risks of preterm delivery, growth restriction, and stillbirth.16 In addition to the effectiveness of vaccination in terms of reducing severe complications, the transplacental transfer of immunoglobulins after maternal vaccination might provide infants with protection against COVID-19.10 16 17 18

Vaccine hesitancy is a potential public health problem.19 The five psychological antecedents of vaccination are confidence, complacency, constraints, calculation, and collective responsibility.20 21 Acceptance of COVID-19 vaccination has varied among countries, with the highest rates reported in India, the Philippines, and Latin America.22 In the present study, Chinese women had a lower vaccination rate, compared with their non-Chinese counterparts.

The COVID-19 pandemic has contributed to a decrease in vaccine hesitancy.23 In May 2021, a systemic review showed that an estimated 47% of pregnant women worldwide intended to undergo COVID-19 vaccination.24

Communities generally become more complacent when the number of disease cases is low, suggesting that the perceived risk of disease transmission is minimal. This phenomenon was evident in Hong Kong across several waves of COVID-19 transmission.15 The vaccination rate increased during the fifth wave of COVID-19 when there was an exponential surge in the number of disease cases. A similar pattern was observed in the present study, such that two-thirds of the antenatal vaccination episodes occurred during the fifth wave of COVID-19 in Hong Kong.

Despite this relative surge in vaccination, around two-thirds of women eligible for the third dose did not receive it during pregnancy. This delay poses a major threat because the protective effect of the previous two doses may have dissipated. Importantly, although the rate of vaccination was higher in the antenatal group than the postnatal group, most women in the antenatal group had undergone vaccination before pregnancy. Thus, they may have a higher risk of serious adverse effects from COVID-19 if they become ill in the peripartum period. After exclusion of the small number of pregnant women aged ≤19 years, vaccination rates were consistently low among pregnant women of all ages. Pregnant women may have a higher risk of disease transmission in future waves of COVID-19; at the end of the present study, only one-third of pregnant women in this cohort had received three doses of vaccine.

Lack of confidence has been identified as the main factor consistently associated with lower COVID-19 vaccine hesitancy.3 The implementation of the stringent vaccine pass policy had driven another peak of vaccination in late May 2022. However, nearly half of the women who had received two doses of vaccine were indeed overdue for the third dose, and hence did not fulfil the vaccine pass for entrance to specific premises or could not attend obstetric clinics without undergoing nucleic acid testing. This group of vaccine had received COVID-19 vaccine before but had it mostly before pregnancy. Pregnancy could represent the key hindrance for their vaccination in the antenatal period. Government policy might not be adequate enough for promoting vaccination in pregnant women. Concern about possible harmful side-effects was the top reason for reluctance; confidence in COVID-19 vaccine safety and efficacy was the main predictor of acceptance, particularly in the pregnant population.8 22 In a Japanese cohort, concerns about potential effects on the fetus and breastfeeding were the main reasons for low COVID-19 vaccination acceptance.25 These findings highlight the need to distribute correct information and provide sufficient education to address concerns among women of reproductive age. In particular, antenatal women should receive additional information concerning vaccine safety during pregnancy.

A study in Hong Kong showed that recommendations from the government constituted the strongest factor driving COVID-19 vaccine acceptance.2 Education about the safety and benefit of vaccination is also important.7 Webinars would be useful in efforts to educate the general public. Within the hospital setting, vaccination teams comprising obstetricians and midwives could allay concerns and dispel myths about vaccination among working staff at all levels; this approach could provide useful information for pregnant and breastfeeding women. There is also a need to combat physician hesitancy in recommending COVID-19 vaccination for pregnant women.26 To address this need, the HKCOG revised its recommendations on 3 March 2022 to indicate that women who are planning to become pregnant, are pregnant, or are breastfeeding should undergo COVID-19 vaccination along with the general population.12 A corresponding educational video to promote COVID-19 vaccination was made available on 5 March 2022.12

Possible interventions to promote vaccination in Hong Kong include the provision of vaccines at convenient venues and the involvement of healthcare professionals in information dissemination.27 During the fifth wave of COVID-19, pregnant women were proactively asked to consider COVID-19 vaccination when they attended obstetric clinics. Leaflets were distributed with information about the HKCOG recommendations, as well as community vaccination sites. The establishment of a pathway specifically for pregnant women, which reduced their waiting time in vaccine clinics, also helped to increase the vaccination rate. Moreover, vaccination was provided to women in the maternity wards of some hospitals and women attending antenatal clinics in Maternal and Child Health Centres. All of these measures helped reduce vaccine hesitancy in pregnant women.28

In Hong Kong, a pertussis vaccination programme for pregnant women was launched on 2 July 2020.29 The programme was incorporated into antenatal care, such that all pregnant women received counselling concerning the rationale for vaccination; the vaccine was administered during obstetric follow-up. To facilitate vaccine availability in our hospital, all women were asked to indicate their preference concerning pertussis vaccination during the first antenatal visit. It may be useful to incorporate COVID-19 vaccination into the maternal immunisation programme.

This study had some notable strengths. The results of the present large cohort study provide clinicians and policymakers with key insights concerning the COVID-19 vaccination rate among pregnant women in Hong Kong. The study period was designed to include both antenatal and postnatal periods for a better understanding of vaccination behaviour among women in each period. A real-time collection method was adopted to capture COVID-19 vaccination records from the Clinical Management System, thereby ensuring data reliability.

Nevertheless, this study had some limitations. A small number of women who underwent COVID-19 vaccination outside Hong Kong were not automatically identified in the system; however, they were included in the cohort if their vaccination history had been documented in antenatal records. Because rapid antigen self-tests were acceptable for diagnosis in Hong Kong beginning on 26 February 2022, the number of recorded COVID-19 cases in our cohort might have been lower than the actual number of cases if patients did not report positive COVID-19 test results to our department. Finally, this study did not explore whether the generally more cautious approach of pregnant women, in terms of avoiding all types of diseases, might contribute to a lower disease rate compared with the general public.

The rate of COVID-19 vaccination was low among pregnant and postnatal women in Hong Kong in early 2022. Pregnant women had a high risk of disease transmission because many of them had not received the third dose of COVID-19 vaccine. Urgent measures are needed to promote vaccination among pregnant women before future waves of COVID-19. In particular, women should receive information concerning vaccine safety to avoid unnecessary delays related to pregnancy.

Concept or design: PW Hui, DTY Chan.
Acquisition of data: All authors.
Analysis or interpretation of data: PW Hui.
Drafting of the manuscript: PW Hui, LM Yeung, JKY Ko, THT Lai, MTY Seto.
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.

All authors have disclosed no conflicts of interest.

The authors thank Dr TW Chay, Dr YF Chiu, Ms WK Choi, Dr TY Hui, Dr KH Lam, Dr KS Lee, Dr YH Luk, Dr SK Ma, Dr HS Wang, Dr CCY Wong, Dr CL Wong, Dr LC Wong, and Dr CWK Yan of Department of Obstetrics and Gynaecology, Queen Mary Hospital for their contributions to research data retrieval and entry.

The abstract has been presented online in the Royal College of Obstetricians and Gynaecologists World Congress 2023 (14 June 2013, London, United Kingdom).

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

Ethics approval was obtained from the Institutional Review Board of The University of Hong Kong/Hospital Authority Hong Kong West Cluster (Ref No.: UW 22-205). Informed patient consent has been waived by the Board due to the retrospective nature of the research.

1. MacDonald NE; SAGE Working Group on Vaccine Hesitancy. Vaccine hesitancy: definition, scope and determinants. Vaccine 2015;33:4161-4. Crossref

2. Wong MC, Wong EL, Huang J, et al. Acceptance of the COVID-19 vaccine based on the health belief model: a population-based survey in Hong Kong. Vaccine 2021;39:1148-56. Crossref

3. Xiao J, Cheung JK, Wu P, Ni MY, Cowling BJ, Liao Q. Temporal changes in factors associated with COVID-19 vaccine hesitancy and uptake among adults in Hong Kong: serial cross-sectional surveys. Lancet Reg Health West Pac 2022;23:100441. Crossref

4. Beigi RH, Krubiner C, Jamieson DJ, et al. The need for inclusion of pregnant women in COVID-19 vaccine trials. Vaccine 2021;39:868-70. Crossref

5. Scientific Committee on Emerging and Zoonotic Disease and Scientific Committee on Vaccine Preventable Diseases, Centre for Health Protection, Hong Kong SAR Government. Consensus interim recommendations on the use of COVID-19 vaccines in Hong Kong (as of Jan 7, 2021). Available from: https://www.chp.gov.hk/files/pdf/consensus_interim_recommendations_on_the_use_of_covid19_vaccines_inhk.pdf. Accessed 30 Mar 2022.

6. Tao L, Wang R, Han N, et al. Acceptance of a COVID-19 vaccine and associated factors among pregnant women in China: a multi-center cross-sectional study based on health belief model. Hum Vaccin Immunother 2021;17:2378-88. Crossref

7. Sznajder KK, Kjerulff KH, Wang M, Hwang W, Ramirez SI, Gandhi CK. COVID-19 vaccine acceptance and associated factors among pregnant women in Pennsylvania 2020. Prev Med Rep 2022;26:101713. Crossref

8. Sutton D, D’Alton M, Zhang Y, et al. COVID-19 vaccine acceptance among pregnant, breastfeeding, and nonpregnant reproductive-aged women. Am J Obstet Gynecol MFM 2021;3:100403. Crossref

9. Shimabukuro TT, Kim SY, Myers TR, et al. Preliminary findings of mRNA COVID-19 vaccine safety in pregnant persons. N Engl J Med 2021;384:2273-82. Crossref

10. Halasa NB, Olson SM, Staat MA, et al. Effectiveness of maternal vaccination with mRNA COVID-19 vaccine during pregnancy against COVID-19–associated hospitalization in infants aged <6 months—17 states, July 2021–January 2022. MMWR Morb Mortal Wkly Rep 2022;71:264-70. Crossref

11. Girardi G, Bremer AA. Scientific evidence supporting coronavirus disease 2019 (COVID-19) vaccine efficacy and safety in people planning to conceive or who are pregnant or lactating. Obstet Gynecol 2022;139:3-8. Crossref

12. The Hong Kong College of Obstetricians and Gynaecologists. The Hong Kong College of Obstetricians and Gynaecologists advice on COVID-19 vaccination in pregnant and lactating women (interim; updated on 3rd March 2022). 2022. Available from: http://www.hkcog.org.hk/hkcog/Upload/EditorImage/20220304/20220304134806_9035.pdf. Accessed 30 Mar 2022.

13. Hong Kong SAR Government. FEHD reminds catering premises operators and customers to observe requirements on third stage of Vaccine Pass. 2022. Available from: https://www.info.gov.hk/gia/general/202205/30/P2022053000726.htm?fontSize=1. Accessed 30 May 2022.

14. Hong Kong SAR Government. Government to implement Vaccine Pass arrangement in designated healthcare premises. 2022. Available from: https://www.info.gov.hk/gia/general/202205/21/P2022052100421.htm. Accessed 21 May 2022.

15. Together, We Fight the Virus. The Government of the Hong Kong Special Administrative Region. Available from: https://www.coronavirus.gov.hk/eng/index.html. Accessed 30 Jun 2022.

16. Allotey J, Stallings E, Bonet M, et al. Clinical manifestations, risk factors, and maternal and perinatal outcomes of coronavirus disease 2019 in pregnancy: living systematic review and meta-analysis. BMJ 2020;370:m3320. Crossref

17. Male V. SARS-CoV-2 infection and COVID-19 vaccination in pregnancy. Nat Rev Immunol 2022;22:277-82. Crossref

18. Nir O, Schwartz A, Toussia-Cohen S, et al. Maternal-neonatal transfer of SARS-CoV-2 immunoglobulin G antibodies among parturient women treated with BNT162b2 messenger RNA vaccine during pregnancy. Am J Obstet Gynecol MFM 2022;4:100492. Crossref

19. Shook LL, Kishkovich TP, Edlow AG. Countering COVID-19 vaccine hesitancy in pregnancy: the “4 Cs”. Am J Perinatol 2022;39:1048-54. Crossref

20. Betsch C, Schmid P, Heinemeier D, Korn L, Holtmann C, Böhm R. Beyond confidence: development of a measure assessing the 5C psychological antecedents of vaccination. PLoS One 2018;13:e0208601. Crossref

21. Chau CY. COVID-19 vaccination hesitancy and challenges to mass vaccination. Hong Kong Med J 2021;27:377-9. Crossref

22. Skjefte M, Ngirbabul M, Akeju O, et al. COVID-19 vaccine acceptance among pregnant women and mothers of young children: results of a survey in 16 countries. Eur J Epidemiol 2021;36:197-211. Crossref

23. Gencer H, Özkan S, Vardar O, Serçekuş P. The effects of the COVID-19 pandemic on vaccine decisions in pregnant women. Women Birth 2022;35:317-23. Crossref

24. Shamshirsaz AA, Hessami K, Morain S, et al. Intention to receive COVID-19 vaccine during pregnancy: a systematic review and meta-analysis. Am J Perinatol 2022;39:492-500. Crossref

25. Hosokawa Y, Okawa S, Hori A, et al. The prevalence of COVID-19 vaccination and vaccine hesitancy in pregnant women: an internet-based cross-sectional study in Japan. J Epidemiol 2022;32:188-94. Crossref

26. Chervenak FA, McCullough LB, Grünebaum A. Reversing physician hesitancy to recommend COVID-19 vaccination for pregnant patients. Am J Obstet Gynecol 2022;226:805-12. Crossref

27. Wang K, Wong EL, Cheung AW, et al. Influence of vaccination characteristics on COVID-19 vaccine acceptance among working-age people in Hong Kong, China: a discrete choice experiment. Front Public Health 2021;9:793533. Crossref

28. Iacobucci G. COVID-19 and pregnancy: vaccine hesitancy and how to overcome it. BMJ 2021;375:n2862. Crossref

29. Hospital Authority. Public hospital pertussis vaccination programme for pregnant women. Jun 28, 2020. Available from: https://www.ha.org.hk/haho/ho/cc/pertussis_press_release_en.pdf. Accessed 29 Jun 2023.

Shoulder dystocia is an uncommon obstetric emergency with an incidence that reportedly ranging from 0.58% to 0.7%.1 2 3 It can result in severe perinatal morbidities, including brachial plexus palsies, clavicular fractures, humeral fractures, hypoxic-ischaemic encephalopathy, cerebral palsy, and even mortality soon after birth.1 2 3 4 5 6 Considering the unpredictable and complex nature of shoulder dystocia, many professional bodies have established systematic approaches with routine training simulations and algorithms to improve fetal outcomes in such cases.1 2 7 8 9 The most common approach is represented by the HELPERR mnemonic, which consists of a sequence of manoeuvres including the McRoberts’ manoeuvre, suprapubic pressure, rotational methods, posterior arm delivery, all-fours position, and clavicular fracture.7 Although the McRoberts’ manoeuvre and suprapubic pressure are often the preferred initial manoeuvres, their success rates (56.0%) are lower than that of the rotational methods (62.4%) and posterior arm delivery (86.1%).5 6 10 The Royal College of Obstetricians and Gynaecologists (RCOG) in 20121 and the American College of Obstetricians and Gynecologists in 20172 revised their guidelines to indicate that either posterior arm delivery or rotational methods can be used after an unsuccessful attempt of the McRoberts’ manoeuvre. Furthermore, a randomised controlled trial published in 2015 demonstrated that the induction of labour at 38 weeks in macrosomic pregnancies could reduce the risk of shoulder dystocia, compared with expectant management.11 The overall caesarean rate did not increase when using this approach. Accordingly, early induction of labour has become an option in cases of suspected fetal macrosomia.

Considering changes in the management of macrosomic pregnancies and shoulder dystocia in the past decade, this study compared the incidences of shoulder dystocia, the maternal and fetal characteristics in such cases, and their obstetric management methods and perinatal outcomes between the periods of 2000-2009 and 2010-2019.

This retrospective study was conducted in a university tertiary obstetric unit that provided obstetric services in the New Territories East Cluster of Hong Kong. All consecutive cases of shoulder dystocia reported from 2000 to 2019 inclusive were identified using the hospital’s electronic database.12 Shoulder dystocia was objectively defined as the requirement of an ancillary obstetric manoeuvre following failed delivery of the anterior shoulder after downward fetal neck traction or head-to-body delivery interval (HBDI) >1 minute, as described in previous reports.4 5 13 Multiple pregnancies, vaginal breech deliveries, and known stillbirths before labour were excluded. Our unit protocol for the management of shoulder dystocia followed the RCOG Green-top Guidelines for shoulder dystocia that was published in 2005 and updated in 2012.1 Beginning in 2002, hands-on training in shoulder dystocia relief was routinely conducted using the ALSO (Advanced Life Support in Obstetrics) program7; the PROMPT (Practical Obstetric Multi-Professional Training)8 and the SOPHIE (Safe Obstetric Practice for High risk and Emergency)9 training methods were added in 2011, after the publication of our articles regarding shoulder dystocia.4 5 The McRoberts’ manoeuvre with or without suprapubic pressure was usually the first manoeuvre performed, followed by a rotational manoeuvre or posterior arm delivery if the McRoberts’ manoeuvre was unsuccessful. A midwife was designated to document each event, including the personnel involved, usage and duration of manoeuvres, and delivery times of the fetal head and fetal body. Umbilical cord blood was collected immediately after delivery for blood gas analysis using a Bayer Rapidpoint 400 Blood Gas Analyzer (Bayer HealthCare, Seattle [WA], United States), as described in our previous reports.14 15 Delivery data, including birth weight and perinatal complications, were recorded immediately after delivery by the attending staff, then crosschecked by another staff member. Two to 3 months later, fetal outcomes were subjected to further review and confirmation during postnatal follow-up and monthly audit meetings. All midwives and obstetricians attended annual training sessions regarding the management of shoulder dystocia.

Identified cases and corresponding medical records were reviewed to collect maternal, fetal, and obstetric characteristics. Advanced maternal age was defined as ≥35 years at the estimated date of delivery. Short stature was defined as maternal height <150 cm. Maternal body weights at booking and delivery were recorded to calculate the body mass index (BMI) at each time point. Obesity was defined as BMI >30 kg/m², in accordance with World Health Organization guidelines.16 17 Parity and diabetes mellitus/gestational diabetes mellitus statuses were noted. Obstetric and neonatal characteristics were recorded, including delivery mode, fetal distress, birth weight, HBDI duration, and the use and sequence of manoeuvres. Macrosomia was defined as birth weight ≥4000 g. Neonatal outcomes were also recorded, including the Apgar scores at 1 minute and 5 minutes of life, cord arterial pH, and neonatal complications (eg, subgaleal haemorrhage, hypoxic-ischaemic encephalopathy, brachial plexus injury, clavicular fracture, and humeral fracture).

The incidences of shoulder dystocia were calculated for three groups: all singleton live pregnancies (excluding multiple pregnancies and stillbirths), all singleton live cephalic-presenting pregnancies with spontaneous onset of labour (excluding cases of fetal malpresentation and elective caesarean deliveries), and all singleton live cephalic-presenting pregnancies with successful vaginal delivery (excluding emergency caesarean deliveries).12 The incidences of shoulder dystocia were also calculated for various birth weight ranges.

The incidences of shoulder dystocia, maternal and fetal characteristics in each case, and perinatal outcomes were compared between 2000-2009 and 2010-2019. Analyses were performed using SPSS (Windows version 26.0; IBM Corp, Armonk [NY], United States). The incidences of shoulder dystocia; maternal, obstetric, and fetal characteristics; and neonatal complications related to shoulder dystocia during 2000-2009 and 2010-2019 were analysed using the Chi squared test and Fisher’s exact test for categorical variables, t test for parametric continuous variables, and Mann-Whitney U test for non-parametric continuous variables. The threshold for statistical significance was set at P<0.05.

In total, this study included 242 cases of shoulder dystocia in the study unit. Table 1 shows the incidences of shoulder dystocia in each decade according to type of birth and range of birth weight. The overall incidence of shoulder dystocia among all singleton live pregnancies decreased from 0.23% (134/58 326) during 2000-2009 to 0.16% (108/65 683) during 2010-2019 (P=0.009). Similarly, the incidence of shoulder dystocia among singleton cephalic-presenting pregnancies with spontaneous onset of labour decreased from 0.25% to 0.19% (P=0.031), and the incidence of shoulder dystocia among singleton cephalic-presenting pregnancies with vaginal delivery decreased from 0.29% to 0.21% (P=0.017). The incidences of shoulder dystocia were generally similar across birth weight categories, but the incidence considerably decreased in the 4200-4399 g group.

Table 2 shows the birth weight distribution for all singleton cephalic live pregnancies. The mean birth weight decreased from 3180 ± 472 g during 2000-2009 to 3132 ± 463 g during 2010-2019 (P<0.001). The proportion of babies weighing ≥4000 g was 3.3% during 2000-2009, whereas it was 2.3% during 2010-2019.

Table 3 illustrates the maternal and obstetric characteristics in cases of shoulder dystocia during each decade. There were no statistically significant differences between the two decades in terms of advanced maternal age, maternal age, maternal height, maternal weight at booking and delivery, BMI at booking and delivery, obesity at delivery, or nulliparity. However, the proportion of shoulder dystocia cases involving maternal diabetes increased from 9.0% during 2000-2009 to 19.4% during 2010-2019 (P=0.018). Additionally, the proportion of shoulder dystocia cases involving instrumental delivery decreased from 65.7% to 47.2% (P=0.004), but there was no statistically significant difference in the proportion of deliveries involving fetal distress. Although there was no significant change in birth weight, the proportion of babies with macrosomia among shoulder dystocia cases tended to decrease over time (from 27.6% to 18.5%; P=0.097).

Table 4 shows the success rates of various manoeuvres in terms of alleviating shoulder dystocia. During 2010-2019, 79.6% of babies in cases of shoulder dystocia had HBDI ≤2 minutes; 14.8% and 5.6% of such babies had HBDI of 3-4 minutes and ≥5 minutes, respectively. These proportions were significantly better than the proportions during 2000-2009 (59.0%, 31.3%, and 9.7%, respectively; P=0.003). The success rate of the McRoberts’ manoeuvre in terms of alleviating shoulder dystocia increased from 31.3% to 47.2% (P=0.012) among all vaginal deliveries, which was partially attributed to the increased success rate among instrumental deliveries (from 20.5% to 39.2%; P=0.017). Although the rotational manoeuvre continued to be preferred over posterior arm extraction (77.6% vs. 22.4%) after failure of the McRoberts’ manoeuvre and suprapubic pressure, the success rate of posterior arm extraction increased from 52.9% in 2000-2009 to 92.3% in 2010-2019 (P=0.042). Table 5 shows the neonatal outcomes of shoulder dystocia. There were significant reductions in the rates of low Apgar scores: for an Apgar score <5 at 1 minute of life, the rate decreased from 13.4% to 5.6% (P=0.042); for an Apgar score <7 at 5 minutes of life, the rate decreased from 11.9% to 4.6% (P=0.045). There were no statistically significant changes in the rates of other neonatal complications.

This study revealed a significant reduction in the overall incidence of shoulder dystocia over the past two decades in a tertiary obstetric unit in Hong Kong. One possible reason is the increased use of caesarean delivery, in both elective and emergency settings, in cases of suspected macrosomia. This hypothesis is supported by the decrease in the incidences of shoulder dystocia among all births and among all pregnancies with onset of labour in the 4200-4399 g subgroup during 2010-2019 (Table 1). However, such decreases were not observed in other subgroups with babies weighing ≥4000 g (4000-4199 g, 4400-4599 g, and ≥4600 g) [Table 1]. These findings suggest that caesarean delivery in cases of suspected macrosomia is not a major factor contributing to shoulder dystocia prevention. Furthermore, the absence of any change in shoulder dystocia incidence among subgroups with babies weighing <4200 g was consistent with our departmental practice of using 4200 g, rather than 4000 g, as a threshold for offering caesarean delivery to non-diabetic women. The use of a lower fetal weight threshold (eg, 4000 g) in pregnant Chinese women could lead to an unnecessary increase in the rate of caesarean delivery.18 The practice of early induction of labour in cases of suspected macrosomia may be the main factor contributing to the decrease in shoulder dystocia incidence.11 Since 2011, women with a fetal abdominal circumference or estimated fetal weight above the 97th percentile (but <4200 g) have been counselled about the risk of difficult labour, along with the benefits and risks of inducing labour to prevent further macrosomia. These approaches are consistent with the decreases in overall mean birth weight and proportion of babies ≥4000 g from 2000-2009 to 2010-2019 (Table 2).

Additionally, our study revealed improvements in the acute management of shoulder dystocia. In particular, the overall success rate of the McRoberts’ manoeuvre in terms of alleviating shoulder dystocia improved from 31.3% during 2000-2009 to 47.2% during 2010-2019; these results were consistent with rates in published reports, which have considerably varied from 23% to 70%.5 6 19 20 21 Our improved success rate was mainly attributed to the increased success rate among instrumental deliveries (from 20.5% to 39.2%) [Table 4]. We previously speculated that instrumental delivery increased the risk of shoulder dystocia while reducing the likelihood of McRoberts’ manoeuvre success, presumably related to delayed descent of the shoulders during instrumental delivery.22 The SOPHIE training emphasises that proper performance of the McRoberts’ manoeuvre should result in cephaloid rotation of the mother’s pelvis, manifested by elevation of the mother’s buttocks from the bed. To achieve this goal, the best method for hyperflexion of the mother’s hips involves grasping the back of the mother’s distal thigh and pushing it in the direction of the mother’s head. By leaning in the same direction, the clinician can use their own weight to facilitate hip hyperflexion. A common mistake is holding the mother’s knee and pushing at the foot. The resulting force is reduced and the mother may experience discomfort at the ankle joint (Fig a and b).9

Importantly, this study showed an increasing trend in the utilisation of posterior arm delivery during 2010-2019, which is consistent with the updated practice guidelines from the RCOG and the American College of Obstetricians and Gynecologists during the same period.1 2 The success rate of posterior arm delivery also substantially increased (from 52.9% to 92.3%) [Table 4] and humeral fracture (a complication associated with posterior arm delivery) did not occur in any case during 2010-2019. The improvement in the success and safety may be related to the enhanced training through the SOPHIE course, which emphasises that the clinician should use the correct hand (ie, right hand for a fetus facing the mother’s left side, and vice versa); this hand should be inserted into the vagina with sufficient depth to reach the fetal posterior forearm, and extraction should be conducted by grasping the forearm (rather than the elbow or upper arm). The direction of the extraction should be outward and upward to generate a rotational effect on the shoulders (Fig c and d).9 10

In addition to enhanced clinician skills, improved management of shoulder dystocia may have resulted from a decline in birth weight from 2000-2009 to 2010-2019, although the difference was not statistically significant. Improvements in the success rates of the McRoberts’ manoeuvre and posterior arm delivery led to improved management of shoulder dystocia, represented by a shorter HBDI (Table 4). Among cases of shoulder dystocia, the proportion of babies with an HBDI of ≤2 minutes increased from 59.0% during 2000-2009 to 79.6% during 2010-2019; conversely, the proportion of such babies with HBDI ≥5 minutes decreased from 9.7% to 5.6% (Table 4). Along with the improvement in HBDI, fewer babies had a low Apgar score. However, such improvements did not lead to significant reductions in the incidences of severe fetal acidosis and hypoxic-ischaemic encephalopathy, possibly because of the small sample size. The management of prenatally missed macrosomia with severe shoulder dystocia remains a substantial challenge. Our group recently reported a severe case of shoulder dystocia in which the posterior shoulder was also lodged in the middle of the pelvic cavity. The situation was resolved by our modified posterior axillary sling traction technique, which involved using a ribbon gauze to form a sling, in combination with long and slim right-angle forceps to facilitate sling placement.23 In our report as well as in our recent review, we emphasised that the sling primarily functions by facilitating shoulder rotation to enable delivery through the wider diagonal diameter of the outlet of the birth canal.10

The strengths of this study include its analysis of a relatively large number of women over two decades, the use of complete and audited outcomes data, and the uniform management of pregnancy complications in accordance with standard guidelines and department protocols.12 However, because this was a retrospective study, causal factors underlying the findings could not be fully elucidated; possible confounding factors included changes in the management of pregnancy complications, management of babies with macrosomia, and management of shoulder dystocia, as well as changes in the levels of skills and experience among clinical personnel during the study period.

More proactive management of macrosomic pregnancies led to decreases in the overall proportion of babies with macrosomia and incidence of shoulder dystocia from 2000-2009 to 2010-2019. Improvements in shoulder dystocia relief skills were demonstrated by increases in the success rates of manoeuvres (eg, the McRoberts’ manoeuvre and posterior arm delivery), as well as decreases in HBDI and Apgar scores.

All authors contributed to the concept or design of the study, acquisition of the data, analysis or interpretation of the data, drafting of the manuscript, and critical revision of the manuscript for important intellectual content. 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.

The corresponding author is the programme director for the Safe Obstetric Practice for High risk and Emergency (SOPHIE) course. Other authors have disclosed no conflicts of interest.

The authors thank the creators of the SOPHIE course (https://www.obg.cuhk.edu.hk/training-education/sophie-course/) and Ms Catherine Chan at the Department of Obstetrics and Gynaecology, The Chinese University of Hong Kong for contributing to the illustrations.

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

Ethical approval was obtained from the Joint Chinese University of Hong Kong–New Territories East Cluster Clinical Research Ethics Committee (Ref No.: CRE 2017.442). Informed patient consent was waived by the Committee due to the retrospective nature of the study.

1. Royal College of Obstetricians and Gynaecologists. Shoulder Dystocia. Green-Top Guideline No. 42. 2nd Edition. London: Royal College of Obstetricians and Gynaecologists. 2012. Available from: https://www.rcog.org.uk/media/ewgpnmio/gtg_42.pdf. Accessed 14 Aug 2023.

2. Practice Bulletin No 178: Shoulder Dystocia [editorial]. Obstet Gynecol 2017;129:e123-33. Crossref

3. Leung TY, Chung TK. Severe chronic morbidity of childbirth. Best Pract Clin Res Obstet Gynaecol 2009;23:401-23. Crossref

4. Leung TY, Stuart O, Sahota DS, Suen SS, Lau TK, Lao TT. Head-to-body delivery interval and risk of fetal acidosis and hypoxic ischaemic encephalopathy in shoulder dystocia: a retrospective review. BJOG 2011;118:474-9. Crossref

5. Leung TY, Stuart O, Suen SS, Sahota DS, Lau TK, Lao TT. Comparison of perinatal outcomes of shoulder dystocia alleviated by different type and sequence of manoeuvres: a retrospective review. BJOG 2011;118:985-90. Crossref

6. Hoffman MK, Bailit JL, Branch DW, et al. A comparison of obstetric maneuvers for the acute management of shoulder dystocia. Obstet Gynecol 2011;117:1272-8. Crossref

7. American Academy of Family Physicians. Advanced Life Support in Obstetrics (ALSO®). Available from: https://www.aafp.org/cme/programs/also.html. Accessed 7 Jan 2022.

8. PROMPT Maternity Foundation. Practical Obstetric Multi-Professional Training. Available from: https://www.promptmaternity.org. Accessed 7 Jan 2022.

9. Department of Obstetrics and Gynaecology, The Chinese University of Hong Kong & Prince of Wales Hospital. SOPHIE Course–Safe Obstetric Practice for High risk and Emergency Course. Available from: https://www.obg.cuhk.edu.hk/training-education/sophie-course/. Accessed 14 Aug 2023.

10. Lau SL, Sin WT, Wong L, Lee NM, Hui SY, Leung TY. A critical evaluation of the external and internal maneuvers for resolution of shoulder dystocia. Am J Obstet Gynecol. In press. Crossref

11. Boulvain M, Senat MV, Perrotin F, et al. Induction of labour versus expectant management for large-for-date fetuses: a randomised controlled trial. Lancet 2015;385:2600-5. Crossref

12. Wong ST, Tse WT, Lau SL, Sahota DS, Leung TY. Stillbirth rate in singleton pregnancies: a 20-year retrospective study from a public obstetric unit in Hong Kong. Hong Kong Med J 2022;28:285-93. Crossref

13. Spong CY, Beall M, Rodrigues D, Ross MG. An objective definition of shoulder dystocia: prolonged head-to-body delivery intervals and/or the use of ancillary obstetric maneuvers. Obstet Gynecol 1995;86:433-6. Crossref

14. Leung TY, Lok IH, Tam WH, Leung TN, Lau TK. Deterioration in cord blood gas status during the second stage of labour is more rapid in the second twin than in the first twin. BJOG 2004;111:546-9. Crossref

15. Leung TY, Tam WH, Leung TN, Lok IH, Lau TK. Effect of twin-to-twin delivery interval on umbilical cord blood gas in the second twins. BJOG 2002;109:63-7. Crossref

16. WHO Expert Consultation. Appropriate body-mass index for Asian populations and its implications for policy and intervention strategies. Lancet 2004;363:157-63. Crossref

17. Leung TY, Leung TN, Sahota DS, et al. Trends in maternal obesity and associated risks of adverse pregnancy outcomes in a population of Chinese women. BJOG 2008;115:1529-37. Crossref

18. Cheng YK, Lao TT, Sahota DS, Leung VK, Leung TY. Use of birth weight threshold for macrosomia to identify fetuses at risk of shoulder dystocia among Chinese populations. Int J Gynaecol Obstet 2013;120:249-53. Crossref

19. McFarland MB, Langer O, Piper JM, Berkus MD. Perinatal outcome and the type and number of maneuvers in shoulder dystocia. Int J Gynaecol Obstet 1996;55:219-24. Crossref

20. Gherman RB, Goodwin TM, Souter I, Neumann K, Ouzounian JG, Paul RH. The McRoberts’ maneuver for the alleviation of shoulder dystocia: how successful is it? Am J Obstet Gynecol 1997;176:656-61. Crossref

21. Spain JE, Frey HA, Tuuli MG, Colvin R, Macones GA, Cahill AG. Neonatal morbidity associated with shoulder dystocia maneuvers. Am J Obstet Gynecol 2015;212:353.e1-5. Crossref

22. Lok ZL, Cheng YK, Leung TY. Predictive factors for the success of McRoberts’ manoeuvre and suprapubic pressure in relieving shoulder dystocia: a cross-sectional study. BMC Pregnancy Childbirth 2016;16:334. Crossref

23. Kwan AH, Hui AS, Lee JH, Leung TY. Intrauterine fetal death followed by shoulder dystocia and birth by modified posterior axillary sling method: a case report. BMC Pregnancy Childbirth 2021;21:672. Crossref