Thyroid nodules are a common clinical finding, with a prevalence of approximately 4% to 7% in the general population, and are most often detected by ultrasonography.1 2 Although most thyroid nodules are benign, distinguishing malignant from benign nodules remains a clinical priority to avoid unnecessary procedures and ensure timely intervention.3 To standardise risk stratification, various Thyroid Imaging Reporting and Data Systems (TIRADS) have been developed,4 5 including the ACR-TIRADS (American College of Radiology),6 the K-TIRADS (Korean Society of Thyroid Radiology),7 and the European Thyroid Association.8 Recognising the need for a system tailored to the Chinese healthcare context, the Chinese Artificial Intelligence Alliance for Thyroid and Breast Ultrasound proposed the Chinese TIRADS (C-TIRADS) in 2021.2 However, existing TIRADS models primarily focus on sonographic characteristics and often overlook relevant clinical risk factors (eg, patient age, sex, and cervical lymph node [LN] involvement).9 In clinical practice, radiologists frequently incorporate such clinical information into their assessments, contributing to inconsistency and variability in TIRADS classification.

Papillary thyroid carcinoma accounts for approximately 80% to 90% of all thyroid cancers and is typically characterised by indolent behaviour.10 11 A substantial proportion of new cases involve papillary thyroid microcarcinoma, defined as tumours measuring less than 10 mm in diameter, which generally carry a favourable clinical prognosis.12 Increasing recognition of the indolent nature of papillary thyroid microcarcinoma has raised concerns regarding potential overdiagnosis and overtreatment. However, current risk stratification strategies that rely solely on imaging features may either overestimate or underestimate malignancy risk, depending on the patient’s broader clinical context. Approaches that incorporate clinical risk factors into TIRADS classification could address these limitations and enhance diagnostic accuracy, supporting more individualised patient management.

This study aimed to develop and externally validate a predictive model that integrates both imaging characteristics and clinical risk factors to refine the C-TIRADS classification system. To our knowledge, this is the first nomogram-based model to incorporate clinical risk factors into the C-TIRADS framework. The tool is designed to assist radiologists in improving diagnostic consistency and supporting more informed and individualised clinical decision making in the management of thyroid nodules.

This retrospective diagnostic study included patients with thyroid nodules who underwent surgical resection at two tertiary hospitals in China. The derivation cohort comprised patients treated at Sichuan Provincial People’s Hospital from January to December 2022, while the external validation cohort was drawn from Affiliated Hospital of North Sichuan Medical College during the same period. Inclusion criteria were: (1) thyroid nodules confirmed by postoperative pathology and (2) preoperative ultrasonography of the thyroid and cervical LNs with complete imaging and clinical records. Exclusion criteria were: (1) unclear pathological diagnosis; (2) incomplete clinical data; or (3) poor-quality ultrasound images.

Two junior radiologists, blinded to clinical and pathological information, independently classified all nodules according to the C-TIRADS criteria. Subsequently, two senior radiologists re-evaluated the cases and adjusted the classifications based on additional clinical risk factors, including patient demographics and cervical LN findings. Any modification from the initial C-TIRADS classification was defined as ‘classification optimisation’ (^*C-TIRADS), encompassing both upgrades and downgrades.

Structured data collection forms were used to record clinical and sonographic variables. The collected data included patient sex, age, nodule size, number of nodules, C-TIRADS classification, and the presence of abnormal cervical LNs on ultrasonography.

Sonographic features that directly determine the C-TIRADS score (such as solidity, echogenicity, aspect ratio, microcalcification, and margin irregularity) were not included independently in the multivariable analysis to avoid collinearity. Based on clinical relevance and univariate regression analysis, six predictors were selected for model development, namely, patient sex, age-group (≤40, 40-60, and >60 years),13 14 nodule size, number of nodules (single vs multiple), presence of abnormal cervical LNs, and original C-TIRADS classification.

A binary logistic regression model was developed using the derivation cohort from Sichuan Provincial People’s Hospital (n=909). For categorical variables with more than two levels, dummy variables were created. The C-TIRADS category 5 was used as the reference group as it represents the highest level of suspicion and the most definitive management pathway (surgical resection), making it an appropriate clinical baseline to estimate relative malignancy risk and the need for reclassification. Model performance in the derivation cohort was evaluated using the area under the receiver operating characteristic (ROC) curve (AUC), and calibration was assessed by comparing predicted probability (PP) with observed outcomes using calibration plots.

We emphasise that the primary outcome variable for model training was the pathological diagnosis (binary: malignant vs benign). The C-TIRADS optimisation, defined as upgrading or downgrading the original category based on PP thresholds, was a post-model clinical decision rule applied to the model output, not the outcome used for model development.

Internal validation was performed using bootstrap resampling with 1000 samples to obtain bias-corrected estimates of model performance and 95% confidence intervals (95% CIs). A fixed random seed was set to ensure reproducibility. The bias-corrected C-statistic was 0.728, compared with the original apparent performance of 0.730 (a difference of 0.002), confirming the model’s stable discriminative ability (online supplementary Table 1).

The final model was applied to the external cohort from Affiliated Hospital of North Sichuan Medical College (n=750) to evaluate its generalisability. Model discrimination was evaluated by calculating the AUC in the validation set, and calibration was assessed using calibration curves.

A nomogram was developed based on the final multivariable regression model to provide a visual tool for clinical application. Each predictor was assigned a score, and the total score corresponded to the PP of C-TIRADS classification optimisation.

Decision curve analysis and clinical impact curves were used to evaluate the clinical utility of the nomogram by quantifying the net benefit across a range of threshold probabilities. Specifically, the nomogram generates a PP indicating whether a nodule’s original C-TIRADS classification should be modified after integrating clinical information. For clinical decision making, we pre-specified probability cut-offs: PP ≥60% (upgrade), PP <30% (downgrade), and PP ≥30% but <60% (unchanged). Based on these thresholds, the model’s recommendations were translated into optimised C-TIRADS categories, which were then compared with radiologists’ optimisation decisions and surgical pathology findings, as appropriate. These thresholds are reported in the Results section and were applied consistently across all performance tables

To ensure consistent ROC analysis, all AUCs were calculated using continuous PPs rather than ordinal risk categories. For the original C-TIRADS system, the five-level ordinal classification was transformed into a continuous malignancy probability score using proportional-odds (ordinal logistic) regression. This standard statistical method was employed to model the ordered nature of the C-TIRADS categories and to derive a continuous probability of malignancy for each category, enabling fair comparison in ROC analysis against other models. For the optimised ^*C-TIRADS system, PPs were directly obtained from the final multivariable logistic regression model. The ROC curves and corresponding AUCs were constructed using these continuous predictions.

Statistical analyses and data visualisation were performed using SPSS (Windows version 26.0; IBM Corp, Armonk [NY], United States) and RStudio (version 2022). Categorical variables were reported as number of cases or percentages, with group comparisons conducted using Chi squared test or Fisher’s exact test, as appropriate. Multivariable logistic regression analysis was conducted to identify independent predictors. Model discrimination was evaluated using ROC curves, while calibration curves were used to assess model accuracy. Clinical decision and impact curves were established to assess practical clinical utility. A two-tailed P value of <0.05 was considered statistically significant.

All models were trained to predict pathological malignancy. The optimised ^*C-TIRADS classifications presented here were derived by applying predefined probability thresholds to the model’s malignancy predictions.

A total of 1659 patients with thyroid nodules were included in the study, comprising 909 patients in the derivation cohort and 750 in the external validation cohort. In the derivation cohort, 71.8% of patients were women, and the majority (90.8%) had nodules measuring ≤30 mm. Approximately 81.7% of patients showed no abnormal cervical LNs on ultrasonography. The rate of C-TIRADS optimisation was 60.6%. In the external validation cohort, similar distributions were observed, with a higher proportion of nodules >30 mm (Table 1).

Univariate binary regression analysis revealed that several variables were either significantly associated (P<0.05) or showed a trend towards association (0.05 < P < 0.1) with C-TIRADS optimisation. These variables included patient sex, age, nodule size (10-30 mm), number of nodules, solid composition, blurred margins, aspect ratio >1, abnormal cervical LNs, and C-TIRADS category (Table 2 and online supplementary Table 2).

A multivariable binary logistic regression model was developed to identify independent predictors associated with C-TIRADS optimisation. Six predictors were independently associated with the outcome. The key predictors of C-TIRADS optimisation were male sex, age 40 to 60 years, thyroid nodule size (per 1-mm increase), multiple thyroid nodules, presence of abnormal cervical LNs, and original C-TIRADS 4A category (online supplementary Table 3). A nomogram model was constructed based on these six independent predictors (Fig 1).

The model demonstrated good discrimination, with an AUC of 0.730 (95% CI=0.697-0.762) in the derivation cohort (online supplementary Fig a). Internal validation using 1000 bootstrap samples yielded a bias-corrected C-statistic of 0.728, indicating stable model performance (online supplementary Table 1). Calibration curves showed good agreement between PPs and observed outcomes (online supplementary Fig b).

Diagnostic thresholds were evaluated to stratify risk. A PP of ≥60% or <30% was considered indicative of a high likelihood of classification change: a PP of ≥60% suggested upgrading, while a PP of <30% suggested downgrading; PPs between 30% and 60% indicated that the classification was likely to remain unchanged. A detailed summary of sensitivity, specificity, and overall accuracy across these thresholds is presented in online supplementary Table 4.

When applied to the external cohort, the model achieved an AUC of 0.865 (95% CI=0.839-0.891) [online supplementary Fig c], demonstrating excellent generalisability. Calibration plots again confirmed close agreement between predicted and observed probabilities (online supplementary Fig d). At the 60% probability threshold, sensitivity was 85.0%, specificity was 69.0%, and overall accuracy was 79.7% in the external validation cohort. Diagnostic performance metrics across various risk thresholds of the final prediction model were analysed in the external validation population (online supplementary Table 5).

Decision curve analysis (Fig 2a) demonstrated that the nomogram model provided greater net clinical benefit across a wide range of threshold probabilities compared with treating all or no patients. The clinical impact curve (Fig 2b) showed that the number of true positives closely approximated the predicted number across relevant thresholds. The observed distribution of histopathological outcomes was as follows: in the derivation cohort, 769 nodules (84.6%) were confirmed malignant and 140 (15.4%) were benign; in the validation cohort, 434 nodules (57.9%) were malignant and 316 (42.1%) were benign.

Comparison of diagnostic efficacy between the original C-TIRADS and optimised C-TIRADS classifications demonstrated superior performance of the optimised model in both the derivation and validation cohorts (Fig 2c and d, respectively). The optimised classification achieved higher AUC values for differentiating benign from malignant nodules (AUC=0.97 vs 0.94 in the derivation cohort; AUC=0.97 vs 0.95 in the external validation cohort). The predictive model tended to improve C-TIRADS classification by upgrading category 4A nodules to category 4B or 4C, reflecting enhanced clinical utility (Table 3 and Fig 2).

A 55-year-old man underwent ultrasound examination, which revealed a solid hypoechoic thyroid nodule in the right lobe measuring approximately 7.1 × 6.4 mm² (Fig 3a). Simultaneously, abnormal LNs were detected on the ipsilateral side of the neck, characterised by indistinct corticomedullary differentiation and suspected microcalcifications (Fig 3b). According to the conventional C-TIRADS system, the nodule was initially classified as category 4B. However, application of the nomogram model yielded a cumulative score of 155 points, corresponding to a malignancy risk of >90%. Based on this result, the TIRADS category was optimised and upgraded to category 5 (Fig 3c). Subsequent histopathological examination confirmed the diagnosis of papillary thyroid microcarcinoma with cervical LN metastasis.

This study retrospectively analysed the sonographic characteristics and clinical risk factors of 1659 thyroid nodules from two large tertiary hospitals in western China, with the aim of optimising the C-TIRADS classification. A predictive model integrating clinical parameters and imaging features was developed and externally validated, demonstrating high diagnostic performance (AUC=0.865 in external validation) and clinical benefit, as evidenced by decision curve analysis.

Despite the widespread adoption of various TIRADS frameworks globally,2 4 5 6 7 8 fundamental methodological limitations persist. Current models, such as ACR-TIRADS,6 primarily focus on ultrasound features and rely heavily on consensus-driven rather than statistically validated risk stratification systems.6 15 Although TIRADS demonstrates robust sensitivity in clinical settings, its specificity remains relatively limited.16 Interobserver variability is another key concern—radiologists’ subjective interpretation of ultrasound features can result in inconsistent classification outcomes.17 To address these limitations, various strategies have been proposed, including the integration of artificial intelligence techniques to reduce observer subjectivity.18 19 20 Artificial intelligence has shown promise in matching or even surpassing the specificity achieved by radiologists; however, their clinical implementation remains constrained by challenges in interpretability and low acceptance in routine practice.

Integrating clinical risk factors may enhance risk stratification for thyroid nodules, as suggested by a growing body of evidence.21 In alignment with this, our study incorporated clinical variables including patient age, sex, number of nodules, and cervical LN status into the predictive model, thereby more accurately reflecting routine clinical diagnostic workflows. While previous studies22 23 24 suggested that male patients with thyroid nodules, particularly those with indeterminate fine-needle aspiration cytology undergoing molecular testing, exhibit a higher malignancy risk,25 our study did not identify a significant difference in thyroid cancer incidence between sexes. This discrepancy may be attributable to methodology differences, as molecular testing was not performed in our cohort and all diagnoses were confirmed through postoperative histopathology. The absence of statistical significance for male sex may reflect population-specific characteristics, such as regional variation in risk factor distribution or age composition.26 These methodological and demographic differences may have attenuated the observed sex-related effect. Nonetheless, male patients in our study were assigned higher risk scores, suggesting an association with malignancy risk, despite the lack of statistical significance.

Compared with previous models that primarily focused on intrinsic ultrasound features of thyroid nodules,27 28 29 our nomogram offers a more comprehensive assessment. Although the individual contributions of factors such as sex and age were relatively modest, they reflected subtle clinical patterns often considered by radiologists during decision making. The C-TIRADS optimisation approach demonstrated clear advantages, particularly in reducing unnecessary invasive procedures without compromising diagnostic accuracy, achieving an AUC of 0.972. Furthermore, the new model indicated that a risk threshold of ≥60% favoured the recommendation for C-TIRADS optimisation, whereas a threshold of <30% favoured exclusion. The integration of complex imaging data with clinical information represents a core competency for radiologists.30 With appropriate standardised training and communication frameworks in place, radiologists are well positioned to leverage quantitative metrics generated by the new model into routine diagnostic workflows. This advancement holds promise for improving diagnostic consistency and accuracy in clinical practice.

This study has several limitations that should be acknowledged. First, the optimisation of the TIRADS classification was influenced by radiologists’ subjective judgement, which may have contributed to interobserver variability. Second, although data collection was conducted by trained junior radiologists, observer variation and the subjective nature of ultrasound interpretation may have affected the model’s performance.31 Third, internal validation using bootstrap resampling may have overestimated model performance due to potential overfitting; therefore, external validation was essential to confirm generalisability. Fourth, owing to the retrospective design, only a limited set of clinical parameters (eg, sex, age, and cervical LN status) was included. Other relevant factors such as body mass index, environmental exposures, nodule location, family history of thyroid cancer, and radiation exposure history,32 33 were not assessed. Finally, the study cohort exclusively comprised cases confirmed by surgical pathology, resulting in a relatively low proportion of benign lesions, which may have introduced selection bias. The exclusion of patients diagnosed solely by fine-needle aspiration was intentional but may have affected the generalisability of the findings.

To address the limitations of the present study, future research should aim to standardise the application of TIRADS by adopting unified classification frameworks and implementing regular training programmes to enhance interobserver consistency. Prospective multicentre studies involving broader and more diverse populations are warranted, incorporating a wider range of clinical risk factors to improve predictive accuracy. In particular, data regarding family history, radiation exposure, and other relevant variables across centres would support more comprehensive risk assessment and enhance the generalisability of prediction models. In addition, including patients with fine-needle aspiration–confirmed benign nodules may help achieve a more balanced representation of benign and malignant cases. The development and application of nomogram-based structured training programmes for radiologists could also be explored to further improve diagnostic consistency and clinical utility. While the widespread adoption of a revised classification system will require time, we hope that the findings of this study may contribute to that transition.

We developed and externally validated a nomogram-based predictive model that integrates imaging features and clinical risk factors to optimise C-TIRADS classification for thyroid nodules. The model demonstrated good discrimination and calibration across internal and external cohorts, offering a practical tool to assist radiologists in refining diagnostic assessments and improving clinical decision making. Future research incorporating additional clinical variables and prospective validation is warranted to further strengthen the model’s applicability across diverse clinical settings.

Concept or design: Y Liang, Y Zou, P He, Q Chen.
Acquisition of data: Y Liang, Y Zou, Z Zou, B Ren.
Analysis or interpretation of data: Y Liang, S Peng, Y Zou.
Drafting of the manuscript: Y Liang, Y Zou, HM Yuan, Z Zou.
Critical revision of the manuscript for important intellectual content: P He, Y Zou.

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 authors have disclosed no conflicts of interest.

This manuscript was initially posted as a preprint entitled ‘Development and validation of a clinical prediction model to aid radiologists optimize thyroid C-TIRADS classification’ on Research Square (DOI: 10.21203/rs.3.rs-3831900/v1). After peer feedback and extensive revisions undertaken collaboratively by the author team, the current version has substantially evolved and markedly differs from the preprint version.

This research was supported by Sichuan Science and Technology Program (Ref Nos.:2025ZNSFSC1751, 2026YFHZ0039), the University-Industry Collaborative Education Program (Ref No.: 250505236300920), the University-level Project of North Sichuan Medical College (Ref Nos.: CXSY24-06, CBY22-QNA48), and the Hospital-level Projects of the Affiliated Hospital of North Sichuan Medical College, China (Ref Nos.: 210930, 2023-2GC013, 2025LC010). The funders had no role in the study design, data collection/analysis/interpretation, or manuscript preparation.

This research was approved by the Ethics Committee of Sichuan Provincial People’s Hospital (Ref No.: ER20210347) and the Ethics Committee of Affiliated Hospital of North Sichuan Medical College, China (Ref No.: 2021ER436-1). The requirement for informed patient consent was waived by both Committees due to the retrospective nature of the research.

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

1. Haugen BR, Alexander EK, Bible KC, et al. 2015 American Thyroid Association management guidelines for adult patients with thyroid nodules and differentiated thyroid cancer: the American Thyroid Association guidelines task force on thyroid nodules and differentiated thyroid cancer. Thyroid 2016;26:1-133. Crossref

2. Zhou J, Song Y, Zhan W, et al. Thyroid imaging reporting and data system (TIRADS) for ultrasound features of nodules: multicentric retrospective study in China. Endocrine 2021;72:157-70. Crossref

3. Trimboli P. Complexity in the interpretation and application of multiple guidelines for thyroid nodules: the need for coordinated recommendations for “small” lesions. Rev Endocr Metab Disord 2025;26:223-7. Crossref

4. Park JY, Lee HJ, Jang HW, et al. A proposal for a thyroid imaging reporting and data system for ultrasound features of thyroid carcinoma. Thyroid 2009;19:1257-64. Crossref

5. Horvath E, Majlis S, Rossi R, et al. An ultrasonogram reporting system for thyroid nodules stratifying cancer risk for clinical management. J Clin Endocrinol Metab 2009;94:1748-51. Crossref

6. Tessler FN, Middleton WD, Grant EG, et al. ACR Thyroid Imaging, Reporting and Data System (TI-RADS): white paper of the ACR TI-RADS Committee. J Am Coll Radiol 2017;14:587-95. Crossref

7. Shin JH, Baek JH, Chung J, et al. Ultrasonography diagnosis and imaging-based management of thyroid nodules: revised Korean Society of Thyroid Radiology consensus statement and recommendations. Korean J Radiol 2016;17:370-95. Crossref

8. Russ G, Bonnema SJ, Erdogan MF, Durante C, Ngu R, Leenhardt L. European Thyroid Association guidelines for ultrasound malignancy risk stratification of thyroid nodules in adults: the EU-TIRADS. Eur Thyroid J 2017;6:225-37. Crossref

9. Chen Z, Wang JJ, Du JB, et al. Development and validation of a dynamic nomogram for predicting central lymph node metastasis in papillary thyroid carcinoma patients based on clinical and ultrasound features. Quant Imaging Med Surg 2025;15:1555-70. Crossref

10. Boucai L, Zafereo M, Cabanillas ME. Thyroid cancer: a review. JAMA 2024;331:425-35. Crossref

11. Zhang J, Xu S. High aggressiveness of papillary thyroid cancer: from clinical evidence to regulatory cellular networks. Cell Death Discov 2024;10:378. Crossref

12. Ma T, Semsarian CR, Barratt A, et al. Rethinking low-risk papillary thyroid cancers <1 cm (papillary microcarcinomas): an evidence review for recalibrating diagnostic thresholds and/or alternative labels. Thyroid 2021;31:1626-38. Crossref

13. Kwong N, Medici M, Angell TE, et al. The influence of patient age on thyroid nodule formation, multinodularity, and thyroid cancer risk. J Clin Endocrinol Metab 2015;100:4434-40. Crossref

14. Pizzato M, Li M, Vignat J, et al. The epidemiological landscape of thyroid cancer worldwide: GLOBOCAN estimates for incidence and mortality rates in 2020. Lancet Diabetes Endocrinol 2022;10:264-72. Crossref

15. Tessler FN, Middleton WD, Grant EG, Hoang JK. Re: ACR Thyroid Imaging, Reporting and Data System (TI-RADS): white paper of the ACR TI-RADS Committee. J Am Coll Radiol 2018;15(3 Pt A):381-2. Crossref

16. Angelopoulos N, Goulis DG, Chrisogonidis I, et al. Diagnostic performance of European and American College of Radiology Thyroid Imaging Reporting and Data System classification systems in thyroid nodules over 20 mm in diameter. Endocr Pract 2025;31:72-9. Crossref

17. Jin Z, Pei S, Shen H, et al. Comparative study of C-TIRADS, ACR-TIRADS, and EU-TIRADS for diagnosis and management of thyroid nodules. Acad Radiol 2023;30:2181-91. Crossref

18. Wildman-Tobriner B, Buda M, Hoang JK, et al. Using artificial intelligence to revise ACR TI-RADS risk stratification of thyroid nodules: diagnostic accuracy and utility. Radiology 2019;292:112-9. Crossref

19. Wu SH, Li MD, Tong WJ, et al. Adaptive dual-task deep learning for automated thyroid cancer triaging at screening US. Radiol Artif Intell 2025;7:e240271. Crossref

20. Trimboli P, Colombo A, Gamarra E, Ruinelli L, Leoncini A. Performance of computer scientists in the assessment of thyroid nodules using TIRADS lexicons. J Endocrinol Invest 2025;48:877-83. Crossref

21. Kobaly K, Kim CS, Mandel SJ. Contemporary management of thyroid nodules. Annu Rev Med 2022;73:517-28. Crossref

22. Xu L, Li G, Wei Q, El-Naggar AK, Sturgis EM. Family history of cancer and risk of sporadic differentiated thyroid carcinoma. Cancer 2012;118:1228-35. Crossref

23. Iglesias ML, Schmidt A, Ghuzlan AA, et al. Radiation exposure and thyroid cancer: a review. Arch Endocrinol Metab 2017;61:180-7. Crossref

24. Saenko V, Mitsutake N. Radiation-related thyroid cancer. Endocr Rev 2024;45:1-29. Crossref

25. Figge JJ, Gooding WE, Steward DL, et al. Do ultrasound patterns and clinical parameters inform the probability of thyroid cancer predicted by molecular testing in nodules with indeterminate cytology? Thyroid 2021;31:1673-82. Crossref

26. Li X, Xing M, Tu P, et al. Urinary iodine levels and thyroid disorder prevalence in the adult population of China: a large-scale population-based cross-sectional study. Sci Rep 2025;15:14273. Crossref

27. Xiao J, Xiao Q, Cong W, et al. Discriminating malignancy in thyroid nodules: the nomogram versus the Kwak and ACR TI-RADS. Otolaryngol Head Neck Surg 2020;163:1156-65. Crossref

28. Xin Y, Liu F, Shi Y, Yan X, Liu L, Zhu J. A scoring system for assessing the risk of malignant partially cystic thyroid nodules based on ultrasound features. Front Oncol 2021;11:731779. Crossref

29. Zhou T, Hu T, Ni Z, et al. Comparative analysis of machine learning-based ultrasound radiomics in predicting malignancy of partially cystic thyroid nodules. Endocrine 2024;83:118-26. Crossref

30. Bluethgen C, Van Veen D, Zakka C, et al. Best practices for large language models in radiology. Radiology 2025;315:e240528. Crossref

31. He Z, Li Y, Zeng W, et al. Can a computer-aided mass diagnosis model based on perceptive features learned from quantitative mammography radiology reports improve junior radiologists’ diagnosis performance? An observer study. Front Oncol 2021;11:773389. Crossref

32. Kim Y, Roh J, Song DE, et al. Risk factors for posttreatment recurrence in patients with intermediate-risk papillary thyroid carcinoma. Am J Surg 2020;220:642-7. Crossref

33. Zhao J, Wen J, Wang S, Yao J, Liao L, Dong J. Association between adipokines and thyroid carcinoma: a meta-analysis of case-control studies. BMC Cancer 2020;20:788. Crossref

Pleural diseases are common respiratory conditions that often require hospital admission and have shown an increasing incidence.1 2 In the United States, approximately 1.5 million patients experience pleural effusion annually, with most cases attributed to congestive heart failure, pneumonia, and cancer.3 4 A recent multicentre, cross-sectional study in China estimated the prevalence of pleural effusion at 4684 per 1 million Chinese adults.5 In that study, the most common causes were parapneumonic effusion and empyema (25.1%), malignant neoplasms (23.7%), and tuberculosis (12.3%).5 The median hospitalisation cost was ¥15 534.5 (interquartile range, 9447.2-29 000.0).5 Additionally, an increasing trend in admissions for spontaneous pneumothorax has been observed in England, highlighting the prevalence of the disease and its associated healthcare burden.2

Management of pleural diseases involves various diagnostic and therapeutic procedures that extend beyond the pleural space to include the airway and lung parenchyma. Whether closed or open, these procedures substantially contribute to the overall healthcare burden. However, information about pleural diseases and related respiratory procedures in Hong Kong remains limited, highlighting the need for contemporary, population-based epidemiological data.

The Hospital Authority, which provides healthcare services to over 90% of Hong Kong’s population, maintains extensive healthcare databases. These include the Clinical Management System (CMS) and the Clinical Data Analysis and Reporting System (CDARS), which capture a wide range of longitudinal clinical data. Examples include hospital discharge records, diagnosis and procedure codes for each hospitalisation episode, radiological findings, and laboratory parameters, particularly blood and pleural fluid analyses. This comprehensive dataset provides valuable insights into the burden of pleural diseases and accurately represents the local population.

Before analysing diseases and procedures using administrative data, it is essential to validate the accuracy of diagnosis and procedure codes within the healthcare database. These codes are typically entered by attending physicians, interventionists, or surgeons performing the procedures, which suggests a high degree of reliability. However, no prior local validation study has been conducted. Therefore, we aimed to assess whether diagnosis codes for pleural diseases and procedure codes for relevant respiratory procedures are accurately recorded for each hospitalisation episode within the Hospital Authority systems.

This retrospective, observational validation study of diagnosis and procedure codes utilised data from a territory-wide healthcare database in Hong Kong. Clinical data were obtained from CDARS, provided by the Hospital Authority. Hospitalisation episodes with the targeted diagnosis and procedure codes between 1 January 2013 and 31 December 2022 were retrieved from the system. Each observation represented a hospitalisation episode rather than a unique patient, and no patient recruitment was involved.

Diagnosis and procedure codes were defined using the International Classification of Diseases, Ninth Revision, Clinical Modification (ICD-9-CM). The basic format of an ICD-9-CM code consists of three to six digits. The Hospital Authority further extends these codes with additional characters after the decimal point to specify particular diagnoses or procedures within an ICD-9-CM code subgroup (‘subcodes’). These subcodes are displayed in CDARS but are not typically accessible to frontline CMS users. All hospitalisation episodes in acute hospitals with a discharge diagnosis code of pneumothorax (codes starting with 512), pleural effusion (codes starting with 012, 197.2, 220.4, 510, or 511), traumatic pneumothorax or haemothorax (trauma-related pleural events, codes starting with 860), or procedure codes for relevant respiratory procedures (codes starting with 33 or 34) were retrieved, regardless of their position in the coding list. Hospitalisation episodes for patients younger than 18 years or from paediatric departments were excluded from subsequent validation analyses. Uninterrupted hospitalisation episodes following the index episodes, including those in acute or convalescent hospitals with the same diagnosis code of interest, were also excluded, as these may represent duplicate entries for the same clinical event. The remaining hospitalisation episodes after exclusions were grouped as the main cohort.

Manual verification of a proportion of the retrieved diagnosis and procedure codes, down to the subcode level, was conducted to ensure data accuracy. The main cohort was first filtered to include only hospitalisation episodes at the authors’ affiliated institution, Prince of Wales Hospital (PWH), forming the PWH cohort. A maximum of 50 hospitalisation episodes for each diagnosis or procedure code were randomly extracted from the PWH cohort to estimate the true positive predictive values (PPVs) within a 13% margin of error at a 95% confidence interval (95% CI). This precision level was chosen pragmatically to balance statistical rigour with the substantial manual effort required for chart review in this validation study. Prince of Wales Hospital is a tertiary care centre with a complex case mix, encompassing a wide range of pleural diseases and advanced respiratory procedures. Within the PWH cohort, the types of pleural disease (pleural effusion, pneumothorax, and trauma-related pleural events) and their underlying aetiologies (eg, non-tuberculous infection, tuberculosis, and malignancy) were determined through retrospective review of clinical notes, discharge summaries, radiological findings, and blood and pleural fluid analysis results using the CMS. Procedure codes were verified by reviewing procedure records within the corresponding hospitalisation episodes. All cases were independently reviewed by two board-certified respiratory physicians. Discrepancies were resolved through joint case review until consensus was reached. Coding accuracy was expressed as PPVs with 95% CIs. The PPV was calculated by dividing the number of true positives (ie, hospitalisation episodes in the PWH cohort where diagnosis and procedure codes were confirmed by manual verification) by the total number of true positives and false positives (ie, episodes where codes were rejected upon manual review). The 95% CI was calculated using the exact binomial method.

We hypothesised that the PPVs for the accuracy of diagnosis and procedure codes would be equal to or greater than 0.700, a commonly used threshold for successful validation.6 7 8 The primary endpoint was the determination of PPVs for the listed diagnosis and procedure codes. All statistical analyses were performed using Python (version 3.12.6).

A total of 26 757 non-traumatic pneumothorax, 218 018 non-traumatic pleural effusion, and 1269 trauma-related pleural events were retrieved from CDARS between 2013 and 2022. Following the exclusion of paediatric patients and uninterrupted hospitalisation episodes, 20 888 non-traumatic pneumothorax, 199 323 non-traumatic pleural effusion, and 1127 trauma-related pleural events remained in the main cohort. Of these, 2451 (11.7%), 24 938 (12.5%), and 251 (22.3%) diagnosis codes for non-traumatic pneumothorax, non-traumatic pleural effusion, and trauma-related pleural events, respectively, were identified from PWH (Fig). Additionally, 185 154 and 106 450 relevant respiratory procedures with ICD-9-CM codes starting with 33 and 34, respectively, were retrieved. After exclusions, 181 770 and 101 336 procedure codes remained, of which 16 078 (8.8%) and 17 299 (17.1%) procedure codes, respectively, were identified from PWH (Fig). Tables 1, 2, and 3 list the diagnosis codes included in the validation analysis for non-traumatic pneumothorax (Table 1), non-traumatic pleural effusion (Table 2) and trauma-related pleural events (Table 3), while Tables 4 and 5 present the procedure codes starting with ‘33’ and ‘34’, respectively; the breakdown of hospitalisation episodes retrieved using these codes, and the numbers remaining after screening, are also shown.

The overall PPVs (95% CIs) for pneumothorax, pleural effusion, trauma-related pleural events, and all diagnosis codes were 0.853 (0.787-0.904), 0.928 (0.903-0.948), 0.957 (0.907-0.981), and 0.919 (0.898-0.936), respectively. The overall PPVs (95% CIs) for procedure codes starting with 33, starting with 34, and for all procedure codes were 0.932 (0.913-0.948), 0.933 (0.916-0.948), and 0.933 (0.920-0.944), respectively.

The PPVs for diagnosis codes related to pneumothorax, pleural effusion, and trauma-related pleural events were all equal to or greater than 0.700, with ranges of 0.700-1.000, 0.833-1.000, and 0.857-1.000, respectively. The lowest PPV (95% CI) was observed for postoperative pneumothorax (procedure code 512.1.2) at 0.700 (0.560-0.812). The highest PPVs were seen for iatrogenic pneumothorax (procedure code 512.1.0) and postoperative haemothorax (procedure code 511.8.7), both at 1.000, with 95% CIs of 0.933-1.000 and 0.762-1.000, respectively. The reasons for false-positive diagnosis codes are summarised in online supplementary Tables 1 to 3, with inappropriate coding of alternative diseases being the most common cause.

The PPVs for procedure codes starting with 33 ranged from 0.700 to 1.000. Procedure codes starting with 34 met the PPV benchmark, except for 34.04.3 (indwelling pleural catheterisation) and 34.09.3 (drainage of the pleural cavity, open). The reasons for false-positive procedure codes are listed in online supplementary Tables 4 and 5, with inappropriate coding of alternative but similar procedures being the most common cause. The low PPV for procedure code 34.04.3 (indwelling pleural catheterisation) arose from its misuse to represent non-tunnelled pleural catheter insertion, or to document the presence of an indwelling pleural catheter (IPC) inserted during prior hospitalisations. Procedure code 34.09.3 (drainage of the pleural cavity, open) failed to meet the PPV benchmark because it was misused to represent closed pleural drainage by drain insertion, rather than an open procedure.

This study is the first to validate diagnosis and procedure codes for pleural diseases using a healthcare database in Hong Kong. All diagnosis codes for pleural diseases and the majority of procedure codes for relevant respiratory procedures met the PPV benchmark of 0.700 or higher. Only procedure codes 34.04.3 (indwelling pleural catheterisation) and 34.09.3 (drainage of the pleural cavity, open) failed to meet the validation criteria.

In 2008, the Hong Kong Thoracic Society reported the burden of lung disease in Hong Kong using local data from various governmental sources; however, pleural diseases were not included in the report.9 Over the subsequent decade, the incidence rates of individual pleural diseases were studied in Hong Kong. However, these studies were limited in scope as they focused on single pleural diseases (eg, empyema,10 11 12 malignant mesothelioma,13 and spontaneous pneumothorax14) or were restricted to single-centre settings.10 11

There is a pressing need for contemporary, population-based epidemiological data covering various pleural diseases in Hong Kong. A recent local survey highlighted heterogeneous practices in the management of pleural diseases among medical clinicians and reflected a lack of awareness and dedicated service infrastructure for pleural diseases.15 Given the rapid advancements in diagnostic strategies and therapeutic options for pleural diseases,16 an accurate and up-to-date assessment of their clinical burden is crucial. Such data provide a foundation for guiding future research, benchmarking healthcare standards in Hong Kong against those of other countries, informing the allocation of future healthcare resources for pleural diseases, and estimating the workload of healthcare professionals managing these conditions. All such service developments should be based on an accurate estimation of the current burden and projected future demand. The use of existing healthcare databases offers a practical approach; however, relevant diagnosis and procedure codes must first be validated. A similar research pathway was followed by Arnold et al,17 who validated diagnosis codes prior to assessing the epidemiology of pleural empyema in English hospitals.17 18

Nearly all PPVs of the diagnosis and procedure codes studied exceeded the benchmark of 0.700. Notably, PPVs for procedure codes were generally higher than those for diagnosis codes. This is because diagnosis codes can be carried over from previous hospitalisation episodes, enabling attending physicians to select active or inactive diagnosis codes regardless of their relevance to the current episode. In contrast, procedure codes cannot be carried over and must be entered manually to reflect procedures performed during the corresponding hospitalisation episode. This requirement contributes to the higher accuracy for procedure codes.

The PPV for procedure code 34.04.3 (indwelling pleural catheterisation) was unexpectedly low due to misuse. The absence of a specific diagnosis code indicating the presence of an IPC, combined with the inclusion of the term ‘pleural’ in the code description, contributed to its incorrect use, particularly during searches for non-tunnelled pleural catheter insertion. Updated diagnosis codes to indicate the status ‘presence of IPC’, or a new procedure code for ‘pleural fluid drainage using an existing IPC’, would accurately reflect the clinical scenario. Once available, such codes should be validated before any analyses of IPC use in territory-wide healthcare databases. Alternatively, establishing a clinical registry for IPC use could facilitate more accurate tracking of patients with both malignant and benign causes of pleural effusion.

Some diagnosis codes (eg, hydrothorax related to dialysis [511.8.3] and hydrothorax as complication of peritoneal dialysis [551.8.8]) and procedure codes (eg, video-assisted thoracoscopy for haemostasis [34.09.4] and injection into thoracic cavity [34.92.0]) were used in other hospitals but not at PWH; therefore, they could not be validated in this study. Within the PWH cohort, alternative diagnosis or procedure codes were used and validated. However, the number of hospitalisation episodes associated with these codes was small, and their impact would be minimal in a territory-wide healthcare data analysis where similar codes are grouped together.

Duplication of subcodes for similar diagnoses or procedures was also noted. Several diagnoses and procedures were represented by different codes, including:

Hydrothorax related to dialysis (511.8.3) and hydrothorax as complication of peritoneal dialysis (511.8.8);

Fibreoptic bronchoscopy (33.22.0) and bronchoscopy (33.23.0);

Endoscopic ultrasonography of bronchus (33.23.3) and endobronchial ultrasonography (33.23.5);

Closed endoscopic biopsy of bronchus (33.24.0), bronchoscopic biopsy (33.24.1), fibreoptic bronchoscopy with biopsy (33.24.2), and flexible bronchoscopy with biopsy of bronchus (33.24.7);

Lung biopsy via endoscopy (33.27.0), bronchoscopic biopsy under fluoroscopic guidance (33.27.1), and flexible bronchoscopy with biopsy of lung (33.27.2);

Video-assisted thoracoscopy for haemostasis (34.09.4) and video-assisted thoracoscopy, haemostasis (34.21.5); and

Chemical pleurodesis (34.92.1) and pleurodesis, chemical (34.92.2).

Researchers should be reminded to search all relevant diagnosis and procedure codes to minimise the risk of missing data for specific diseases or procedures during code searches. In the long term, reconciling similar codes may help reduce ambiguity and improve data consistency.

This study has several strengths, notably its status as the first validation study conducted using a large healthcare database in Hong Kong. It successfully validated codes for a wide range of pleural diseases and respiratory procedures, thereby laying the foundation for future epidemiological research. However, several limitations should be acknowledged. Not all codes could be adequately validated due to their small case volumes in the PWH cohort. For example, codes for Meigs’ syndrome (220.4), traumatic pneumothorax with open wound into thorax (860.1), and traumatic haemothorax with open wound into thorax (860.3) had small numbers even in the overall cohort, and some codes were duplicated. As such, future research incorporating patient searches based on these diagnosis and procedure codes should take these limitations into account. The single-centre nature of the study represents a further limitation, as disease patterns and coding practices may vary across district general hospitals.

This is the first validation study of diagnosis codes for pleural diseases and procedure codes for relevant respiratory procedures using a territory-wide healthcare database in Hong Kong. All diagnosis codes and the majority of procedure codes demonstrated high PPVs, indicating accurate coding. Given the emergence of new respiratory procedures, diagnosis and procedure codes should be regularly updated. The removal or consolidation of duplicated subcodes within the Hospital Authority system is also necessary to facilitate accurate future research and analysis using clinical codes. Further evaluation and harmonisation of coding practices across different hospitals would be beneficial. These measures will pave the way for future territory-wide studies and enable monitoring of the overall burden of pleural diseases in Hong Kong.

Concept or design: KKP Chan.
Acquisition of data: KKP Chan, TCC Ng, CY Sze, KC Ling.
Analysis or interpretation of data: KKP Chan, TCC Ng, CY Sze, KC Ling.
Drafting of the manuscript: KKP Chan.
Critical revision of the manuscript for important intellectual content: KKP Chan, TCC Ng, C Chan, CHY Lau, SWT Ho, JKC Ng, RLP Lo, WH Yip, JCL Ngai, KW To, FWS Ko, DSC Hui.

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 Prof Terry CF Yip from the Department of Medicine and Therapeutics of The Chinese University of Hong Kong for providing statistical support.

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, Hong Kong (Ref No.: 2022.031). 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. Accepted supplementary material will be published as submitted by the authors, without any editing or formatting. Any opinions or recommendations discussed are solely those of the author(s) and are not endorsed by the Hong Kong Academy of Medicine and the Hong Kong Medical Association. The Hong Kong Academy of Medicine and the Hong Kong Medical Association disclaim all liability and responsibility arising from any reliance placed on the content.

1. Bodtger U, Hallifax RJ. Epidemiology: why is pleural disease becoming more common? In: Maskell NA, Laursen CB, Lee YCG, et al, editors. Pleural Disease. Vol 87. Schweiz, Switzerland: European Respiratory Society; 2020: 1-12. Crossref

2. Hallifax RJ, Goldacre R, Landray MJ, Rahman NM, Goldacre MJ. Trends in the incidence and recurrence of inpatient-treated spontaneous pneumothorax, 1968-2016. JAMA 2018;320:1471-80. Crossref

3. Light RW. Pleural effusions. Med Clin North Am 2011;95:1055-70. Crossref

4. Taghizadeh N, Fortin M, Tremblay A. US hospitalizations for malignant pleural effusions: data from the 2012 National Inpatient Sample. Chest 2017;151:845-54. Crossref

5. Tian P, Qiu R, Wang M, et al. Prevalence, causes, and health care burden of pleural effusions among hospitalized adults in China. JAMA Netw Open 2021;4:e2120306. Crossref

6. Kwok WC, Tam TC, Sing CW, Chan EW, Cheung CL. Validation of diagnostic coding for bronchiectasis in an electronic health record system in Hong Kong. Pharmacoepidemiol Drug Saf 2023;32:1077-82. Crossref

7. Ye Y, Hubbard R, Li GH, et al. Validation of diagnostic coding for interstitial lung diseases in an electronic health record system in Hong Kong. Pharmacoepidemiol Drug Saf 2022;31:519-23. Crossref

8. Kwok WC, Tam TC, Sing CW, Chan EW, Cheung CL. Validation of diagnostic coding for asthma in an electronic health record system in Hong Kong. J Asthma Allergy 2023;16:315-21. Crossref

9. Chan-Yeung M, Lai CK, Chan KS, et al. The burden of lung disease in Hong Kong: a report from the Hong Kong Thoracic Society. Respirology 2008;13 Suppl 4:S133-65. Crossref

10. Chan KP, Ng SS, Ling KC, et al. Phenotyping empyema by pleural fluid culture results and macroscopic appearance: an 8-year retrospective study. ERJ Open Res 2023;9:00534-2022. Crossref

11. Tsang KY, Leung WS, Chan VL, Lin AW, Chu CM. Complicated parapneumonic effusion and empyema thoracis: microbiology and predictors of adverse outcomes. Hong Kong Med J 2007;13:178-86.

12. Chan KP, Ma TF, Sridhar S, Lam DC, Ip MS, Ho PL. Changes in etiology and clinical outcomes of pleural empyema during the COVID-19 pandemic. Microorganisms 2023;11:303. Crossref

13. Chang KC, Leung CC, Tam CM, Yu WC, Hui DS, Lam WK. Malignant mesothelioma in Hong Kong. Respir Med 2006;100:75-82. Crossref

14. Chan JW, Ko FW, Ng CK, et al. Management and prevention of spontaneous pneumothorax using pleurodesis in Hong Kong. Int J Tuberc Lung Dis 2011;15:385-90.

15. Lui MM, Yeung YC, Ngai JC, et al. Implementation of evidence on management of pleural diseases: insights from a territory-wide survey of clinicians in Hong Kong. BMC Pulm Med 2022;22:386. Crossref

16. Lui MM, Lee YC. Twenty-five years of respirology: advances in pleural disease. Respirology 2020;25:38-40. Crossref

17. Arnold DT, Hamilton FW, Morris TT, et al. Epidemiology of pleural empyema in English hospitals and the impact of influenza. Eur Respir J 2021;57:2003546. Crossref

18. Hamilton F, Arnold D. Accuracy of clinical coding of pleural empyema: a validation study. J Eval Clin Pract 2020;26:79-80. Crossref

Hepatocellular carcinoma (HCC) is one of the leading malignancies worldwide. At diagnosis, up to 30% of patients have intermediate-stage HCC according to the Barcelona Clinic Liver Cancer system.1 Transarterial chemoembolisation (TACE) has emerged as the first-line treatment for intermediate-stage HCC, supported by two randomised controlled trials2 3 and a meta-analysis4 that demonstrated superior survival outcomes compared with best supportive care or suboptimal therapies.

Because patients with intermediate-stage HCC comprise a heterogeneous group characterised by a wide range of tumour burdens and liver function, the effectiveness of TACE as first-line treatment may not be universal, particularly in subgroups with high tumour burden or suboptimal liver function. To address this issue, sub-staging of intermediate-stage HCC based on tumour burden and liver function has been proposed in several criteria, including the Bolondi,5 Kinki,6 and MICAN (Modified Intermediate Stage of Liver Cancer) criteria.7 These criteria have demonstrated discriminative prognostic value in identifying subgroups of patients with intermediate-stage HCC.7 8 Given that survival outcomes of patients treated with TACE can vary across substages of intermediate-stage HCC, it is clinically essential to identify thresholds of tumour burden and liver function that preclude the use of TACE according to survival benefit.

Sorafenib has been established as the standard of care for advanced HCC since 2007, based on the demonstration of its significant survival superiority over placebo.9 10 11 Subgroup analyses of clinical trials have shown that sorafenib exerts positive therapeutic efficacy in intermediate-stage HCC, with reported overall survival (OS) ranging from 14.5 to 20.6 months,9 12 13 which is comparable to the OS achieved with TACE. Sorafenib treatment can serve as a benchmark for evaluating the survival benefit of TACE. If TACE does not provide a significant survival benefit compared with sorafenib, it may not be appropriate to subject patients to TACE rather than systemic therapy, given that TACE is invasive and potentially harmful to the liver. Patients may benefit from systemic therapy before liver function becomes suboptimal.

It has been hypothesised that specific baseline parameters of tumour burden and liver function, at which TACE fails to show superior survival benefit compared with sorafenib, could be defined as indicators of TACE unsuitability. This study aimed to define specific indicators of TACE unsuitability at baseline in patients with intermediate-stage HCC according to thresholds of tumour burden and liver function, as judged by the survival benefit of TACE over sorafenib.

Due to the limited number of eligible participants, all available cases with complete clinical data were included. All patients with unresectable HCC who received TACE or sorafenib therapy from January 2005 to December 2019 at Prince of Wales Hospital were enrolled in the study, provided they met all eligibility criteria. Unresectability of intermediate-stage HCC was determined by a multidisciplinary team comprising a surgeon, an interventional radiologist, and an oncologist. Inclusion criteria were treatment-naïve, Barcelona Clinic Liver Cancer-B stage HCC diagnosed by biopsy or a typical vascular pattern on cross-sectional imaging; intrahepatic disease without vascular invasion; and an Eastern Cooperative Oncology Group performance status score of 0 or 1. Exclusion criteria included age under 18 years or Eastern Cooperative Oncology Group performance status score of 2 or above; prior treatment before initial TACE; receipt of hepatectomy, liver transplantation, or local therapy after initial TACE; and any imaging evidence from computed tomography (CT), magnetic resonance imaging, or positron emission tomography/CT showing vascular invasion by tumour (including portal vein tumour thrombus) or extrahepatic metastasis (Fig 1). To identify thresholds for TACE unsuitability, OS of patients treated with TACE was compared with that of patients treated with sorafenib within subgroups defined by baseline tumour burden and liver function. Overall survival was defined as the interval between the initiation of TACE or sorafenib and death from any cause. Patients who were alive or lost to follow-up were censored.

In total, 420 patients were enrolled in the study: 358 received TACE and 62 received sorafenib (Table 1). The TACE group included significantly more older and female patients. The median tumour size was significantly larger in the sorafenib group compared with the TACE group. No significant differences were observed between the two groups in terms of the modified albumin–bilirubin (mALBI) grade distribution or tumour multiplicity. Among patients initially treated with TACE, the median number of TACE sessions was two (range, 1-4); 124 patients received one session, 78 received two sessions, 53 received three sessions, and 103 received more than three sessions. After developing refractoriness to TACE, 60 patients subsequently received systemic agents; of these, 35 received sorafenib, eight received adriamycin, four received doxorubicin, six received lenvatinib, and seven received other agents.

Patients were classified into six subgroups according to baseline tumour burden and liver function. Tumour burden was subcategorised using the up-to-7, up-to-11, and N3-S5-S10 criteria. The up-to-7 and up-to-11 criteria were derived from the sum of the maximum tumour size (in cm) and the tumour number, with cut-off values of 7 or 11, respectively. Accordingly, patients were categorised as within or beyond the up-to-7 and up-to-11 criteria. In the N3-S5-S10 system, tumour burden was subcategorised according to the combination of tumour number and maximum tumour size; three tumour nodules and 5 cm or 10 cm in size served as the respective cut-off values. This categorisation resulted in the following six subgroups: (1) tumour number ≤3, tumour size ≤5 cm; (2) tumour number ≤3, tumour size >5 cm to ≤10 cm; (3) tumour number ≤3, tumour size >10 cm; (4) tumour number >3, tumour size ≤5 cm; (5) tumour number >3, tumour size >5 cm to ≤10 cm; and (6) tumour number >3, tumour size >10 cm (Fig 1).

Liver function subgroups were classified according to the mALBI grade.14 The mALBI grades were determined using the ALBI score, calculated as (log₁₀ [bilirubin level (μmol/L)] × 0.66) + (albumin level [g/L] × –0.085). Based on three cut-off ALBI scores, grades were defined as follows: grade 1 (≤–2.60), grade 2a (>–2.60 to ≤–2.27), grade 2b (>–2.27 to ≤–1.39), and grade 3 (>–1.39). Because the sample size of patients receiving sorafenib with mALBI grade 1 or 2a was relatively small, these two subgroups were combined for analysis. Additionally, given that no patient with mALBI grade 3 received sorafenib, this subgroup was excluded from the analysis (Fig 1).

The TACE procedures were performed using digital subtraction angiography equipment via a femoral approach under local anaesthesia.15 16 In brief, a microcatheter was used to catheterise tumour-feeding arteries at the lobar, segmental, or subsegmental level, depending on tumour size. An emulsion of cisplatin–ethiodised oil (Platosin; Pharmachemie BV, Haarlem, the Netherlands), consisting of up to 20 mg aqueous cisplatin (20 mL) and up to 20-mL ethiodised oil mixed in a 1:1 volume ratio, was administered until flow stasis occurred or a maximum dose of 40-mL emulsion was delivered. Digital subtraction angiography, with or without non-contrast multiplanar CT, was used to confirm treatment completeness. A gelatin sponge (5-10 mL) was used to embolise the feeding arteries.

Postoperative monitoring included blood tests for liver function and tumour markers within 2 days, at 2 weeks, and then every 1 to 3 months, as well as CT imaging every 3 months. Systemic therapy was administered to patients with well-preserved liver function who developed TACE refractoriness, as indicated by continuous elevation of tumour markers and CT evidence of tumour progression.

According to the customary protocol at Prince of Wales Hospital, The Chinese University of Hong Kong during the study period, patients with unresectable intermediate-stage HCC and no contraindications to TACE were prioritised for TACE treatment. Patients who declined TACE were treated with sorafenib; as a result, some patients in the sorafenib group had smaller tumours or fewer tumour nodules. Sorafenib was administered orally at a prescribed dose of 400 mg twice daily. In the event of intolerable side-effects or serious adverse events, oncologists could adjust the treatment by reducing the dose or discontinuing the drug.

Categorical variables were presented as numbers (percentages), while continuous variables were summarised as median (interquartile range), median (95% confidence interval [95% CI]), or depending on the results of normality testing. The Chi squared test was used to compare categorical data, and the Mann-Whitney U test was performed for continuous data. Kaplan-Meier curves and Cox proportional hazards models were used to compare OS values among subgroups. The log-rank test and hazard ratio (HR) were utilised to assess survival differences between subgroups. A sensitivity analysis of survival outcomes was conducted, excluding participants who received systemic therapy after TACE. A P value <0.05 was considered statistically significant. Statistical analyses were performed using SPSS (Windows version 25.0; IBM Corp, Armonk [NY], United States).

The median OS of all patients who received TACE was significantly longer than that of patients who received sorafenib (19.37 [16.89-21.85] months vs 5.12 [4.37-5.84] months, P<0.001; Fig 2a). When stratified by mALBI grade, patients with mALBI grade 1 or 2a had significantly longer median OS in the TACE group compared with the sorafenib group (23.83 [18.53-29.13] months vs 6.60 [3.61-9.59] months, P<0.001; Fig 2b). Similarly, patients with mALBI grade 2b had significantly longer median OS in the TACE group than in the sorafenib group (16.20 [11.91-20.49] months vs 4.39 [3.44-5.35] months, P<0.001; Fig 2c).

The median OS of patients treated with sorafenib, stratified by mALBI grade and tumour burden, is summarised in Table 2. As the sorafenib subgroups with tumour number ≤3 had a relatively small sample size (n=8) according to the N3-S5-S10 criteria, these patients were not further subdivided based on tumour size. Instead, they were combined into a single subgroup with tumour number ≤3 to increase the sample size for comparison with the TACE group. Consequently, OS in the combined sorafenib subgroup (tumour number ≤3, any tumour size) was used for comparison with OS in the three tumour-size TACE subgroups of tumour number ≤3 (Table 2).

The distribution of sample sizes was uneven across the sorafenib subgroups with tumour number >3 based on the N3-S5-S10 criteria, which may have introduced bias in the survival outcomes, such as a lower tumour burden being associated with worse OS. To avoid underestimation of OS in any tumour-size subgroup when comparing with the TACE subgroups, the longest OS among the subgroups with tumour number >3 was utilised as the OS value for all these subgroups in the analysis, irrespective of tumour size (Table 2). As no patients with mALBI grade 2 were present in the tumour burden subgroup defined as within up-to-7, the OS of patients with tumour burden beyond up-to-7 (Table 2) who were treated with sorafenib was used as the control.

Table 3 presents the median OS of patients treated with TACE or sorafenib, stratified by mALBI grade 1 or 2a and tumour burden. Across all subgroups defined by various tumour burden criteria, patients who received TACE achieved significantly longer OS than those who received sorafenib (all P<0.05), with HRs favouring TACE (ranging from 0.130 to 0.331). Sensitivity analysis showed that survival was not significantly different between TACE and sorafenib in the subgroup with tumour number >3 and tumour size >10 cm (HR=0.418 [95% CI=0.147-1.171]; P=0.097).

In subgroups with mALBI grade 2b, defined by either the up-to-7 or up-to-11 criteria, patients who received TACE exhibited significantly longer median OS than those who received sorafenib across all subgroups (all P<0.05; Table 4). However, when using the N3-S5-S10 criteria, TACE resulted in a significantly longer median OS than sorafenib only in the subgroups with tumour number ≤3 (any tumour size) and in the subgroup with tumour number >3 and tumour size ≤5 cm (both P<0.05; Table 4). In the subgroups with tumour number >3 and tumour size >5 cm to ≤10 cm, and those with tumour number >3 and tumour size >10 cm, although TACE subgroups demonstrated longer median OS than sorafenib subgroups (6.07 vs 3.74 months and 7.73 vs 3.74 months, respectively), the differences were not statistically significant (Table 4). Sensitivity analysis showed that survival was also not significantly different between TACE and sorafenib in the additional subgroup with tumour number ≤3 and tumour size >10 cm (HR=0.474 [95% CI=0.185-1.261]; P=0.120).

Due to the small sample size, it was difficult to demonstrate a clear survival benefit of TACE over sorafenib; thus, the risk of overestimating the survival benefit of TACE, due to potential bias from more advanced disease in the sorafenib group, was likely minimised. For example, given the limited number of patients in the subgroups with tumour number >3 and tumour size >5 cm to ≤10 cm and those with tumour number >3 and tumour size >10 cm, these two subgroups were combined into one subgroup (tumour number >3 and tumour size >5 cm). In this combined subgroup, TACE (n=38) still yielded no significant survival benefit over sorafenib (n=14), with OS values of 6.07 months (4.10-8.03) and 3.74 months (1.71-5.78), respectively (HR=0.586 [95% CI=0.325-1.054]; P=0.071).

Subgroup analysis in this study revealed that, within the limitations of the data, TACE probably did not confer a statistically significant survival benefit over sorafenib for patients with mALBI grade 2b and a high tumour burden (number >3 and size >5 cm, or number ≤3 and size >10 cm), or for patients with mALBI grade 1 or 2a and tumour burden of number >3 and size >10 cm. In contrast, TACE did provide a survival benefit when the beyond up-to-7 or beyond up-to-11 criteria were applied. These findings suggest that the use of more precise criteria to define tumour burden and liver function could help identify specific subgroups unsuitable for TACE. Such criteria highlight the threshold at which TACE no longer provides a survival advantage over sorafenib, thereby indicating TACE unsuitability. These indicators would be valuable in guiding the clinical management of intermediate-stage HCC. The small sample size in the sorafenib group may have limited the statistical power to detect a survival benefit of TACE in subgroups with tumour number >3 and size >5 cm. Given that the overall results showed a consistent trend favouring TACE, validation through further studies with larger sample sizes is warranted.

In recent years, systemic therapy for HCC has undergone rapid development, leading to the emergence of new drugs after sorafenib. The combination of certain agents has shown significant improvements in survival compared with sorafenib alone. The IMbrave150 study demonstrated that treatment with atezolizumab plus bevacizumab resulted in a significantly longer median OS than sorafenib alone (19.2 vs 13.4 months).17 Similarly, both sintilimab plus a bevacizumab biosimilar18 and tremelimumab plus durvalumab19 provided significant survival benefits over sorafenib in patients with unresectable HCC. Nevertheless, sorafenib remains the first-line standard treatment and the most effective single agent for advanced HCC. It serves as a benchmark for newer single-agent therapies such as lenvatinib, nivolumab, and durvalumab, which have shown statistical non-inferiority in survival compared with sorafenib.19 20 21 Therefore, the use of sorafenib as the control arm versus TACE in this study is reasonable. With the rapid advancement of systemic agents, novel treatment strategies—such as switching to systemic therapy22 or initiating systemic therapy upfront followed by curative conversion23—have been advocated for patients with intermediate-stage HCC who may not benefit from TACE or repeated TACE. In such cases, it is important to define specific indicators of TACE unsuitability among patients with intermediate-stage HCC, in whom systemic therapy may potentially improve survival.

The concept of TACE unsuitability has emerged in conjunction with the development and availability of systemic therapies.24 In patients with intermediate-stage HCC, TACE unsuitability has been defined as the presence of mALBI grade 2b and tumour burden beyond the up-to-7 criteria.25 26 This definition was based on worse survival in patients with mALBI grade 2b and the beyond up-to-7 criteria relative to patients displaying better liver function and lower tumour burden, without addressing the potential survival benefit of TACE over alternative treatment options in this subgroup. However, this definition has two key limitations. First, it lacks clinical evidence demonstrating greater survival benefit from other alternative treatments when TACE is withheld. Second, there remains controversy regarding the optimal criteria for defining high tumour burden. If the beyond up-to-7 criteria is used as the criterion for TACE unsuitability, the majority of patients with intermediate-stage HCC would be considered unsuitable, which is both unrealistic and unsupported. In the present study, 79% of patients had high tumour burden beyond up-to-7, comparable to the 70% reported by Hung et al.27

The sub-staging system using the up-to-11 criteria has shown better discriminatory power than the up-to-7 criteria for predicting survival after TACE.28 29 Nonetheless, in this study, neither the up-to-7 nor the up-to-11 criteria were able to identify TACE unsuitability. The findings indicated that both the patient subgroup with mALBI grade 2b and tumour burden beyond the up-to-7 criteria, as well as the subgroup with mALBI grade 2b and tumour burden beyond the up-to-11 criteria, still derived survival benefits from TACE compared with sorafenib, indicating that these subgroups should not be considered TACE unsuitable. The lack of discriminatory power may be attributed to the persistently high heterogeneity among patients classified as having high tumour burden under to these two criteria. Worse survival after TACE in these subgroups, compared with patients displaying better liver function and lower tumour burden, does not justify entirely abandoning TACE in these patients.

We propose using the N3-S5-S10 criteria to define tumour burden, as these criteria allow for more specific subgrouping and enable the identification of TACE unsuitability with greater precision, thereby reducing the likelihood of denying patients a potentially beneficial treatment (TACE). Our findings demonstrate that the proposed criteria can identify TACE unsuitability precisely in specific subgroups where the up-to-7 or up-to-11 criteria fail to distinguish survival differences. Based on these findings, we recommend that physicians assess intermediate-stage HCC using both the mALBI grade and the N3-S5-S10 criteria—a more rigorous framework—to determine TACE unsuitability. To our knowledge, this is the first study to demonstrate the survival benefit of TACE over sorafenib in patients with intermediate-stage HCC stratified by both liver function and tumour burden, as well as to identify TACE unsuitability within these subgroups.

This study provided a larger sample size than previous studies comparing survival benefits between TACE and sorafenib. However, several limitations should be noted. First, the retrospective design of this study inevitably introduced patient selection bias between the TACE and sorafenib groups. Although there were significant differences in age, sex, and tumour size between the groups, such disparities in overall patient demographics might not have critically affected the validity of the survival comparisons, given that these were based on subgroup analyses. Second, the sample size was exceedingly small in some sorafenib subgroups with low tumour burden. The substantial disparity in patient numbers may have contributed to non-significant differences in OS between subgroups. We attempted to mitigate this limitation by combining subgroups with very small sample sizes. Third, some patients in the TACE group received systemic therapy after disease progression. Consequently, survival in the TACE group may have been overestimated as it reflected outcomes of TACE with or without systemic therapy, rather than TACE alone. Nonetheless, ‘TACE followed by systemic therapy’ represents standard clinical practice aimed at achieving the greatest patient benefit, and isolating a TACE-alone group for analysis would not be realistic. Notably, ‘TACE followed by systemic therapy’ accurately reflects real-world treatment practice and does not conflict with the study’s primary objective, which was to define specific indicators of TACE unsuitability at baseline rather than at the point when TACE becomes unsuitable. Finally, no power calculation was performed in the statistical analysis.

More precise criteria for TACE unsuitability are required. The combination of mALBI grade and N3-S5-S10 criteria may serve as a better indicator of TACE unsuitability than the beyond up-to-7 or beyond up-to-11 criteria for patients with intermediate-stage HCC. TACE likely offers no survival benefit compared with sorafenib beyond these thresholds. However, validation in a larger cohort is warranted.

Concept or design: SCH Yu.
Acquisition of data: LM Chen, L Li, EP Hui, W Yeo, SL Chan.
Analysis or interpretation of data: LM Chen, SCH Yu.
Drafting of the manuscript: LM Chen, SCH Yu.
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 was funded by the Vascular and Interventional Radiology Foundation, Hong Kong. The funding body was not involved in the design of the study, collection of data, analysis/interpretation of data, or writing of the manuscript.

This research was approved by The Chinese University of Hong Kong–New Territories East Cluster Ethics Committee, Hong Kong (Ref No.: 2020.672). It was conducted in accordance with the Declaration of Helsinki and the International Conference on Harmonisation–Good Clinical Practice guidelines. The requirement for written informed patient consent was waived by the Committee due to the retrospective nature of the research.

1. Sung H, Ferlay J, Siegel RL, et al. Global Cancer Statistics 2020: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA Cancer J Clin 2021;71:209-49. Crossref

2. Llovet JM, Real MI, Montaña X, et al. Arterial embolisation or chemoembolisation versus symptomatic treatment in patients with unresectable hepatocellular carcinoma: a randomised controlled trial. Lancet 2002;359:1734-9. Crossref

3. Lo CM, Ngan H, Tso WK, et al. Randomized controlled trial of transarterial lipiodol chemoembolization for unresectable hepatocellular carcinoma. Hepatology 2002;35:1164-71. Crossref

4. Llovet JM, Bruix J. Systematic review of randomized trials for unresectable hepatocellular carcinoma: chemoembolization improves survival. Hepatology Feb 2003;37:429-42. Crossref

5. Bolondi L, Burroughs A, Dufour JF, et al. Heterogeneity of patients with intermediate (BCLC B) hepatocellular carcinoma: proposal for a subclassification to facilitate treatment decisions. Semin Liver Dis 2012;32:348-59. Crossref

6. Kudo M, Arizumi T, Ueshima K, Sakurai T, Kitano M, Nishida N. Subclassification of BCLC B stage hepatocellular carcinoma and treatment strategies: proposal of modified Bolondi’s subclassification (Kinki criteria). Dig Dis 2015;33:751-8. Crossref

7. Hiraoka A, Kumada T, Nouso K, et al. Proposed new sub-grouping for intermediate-stage hepatocellular carcinoma using albumin–bilirubin grade. Oncology 2016;91:153-61. Crossref

8. Arizumi T, Ueshima K, Iwanishi M, et al. Validation of Kinki criteria, a modified substaging system, in patients with intermediate stage hepatocellular carcinoma. Dig Dis 2016;34:671-8. Crossref

9. Bruix J, Raoul JL, Sherman M, et al. Efficacy and safety of sorafenib in patients with advanced hepatocellular carcinoma: subanalyses of a phase III trial. J Hepatol 2012;57:821-9. Crossref

10. Cheng AL, Kang YK, Chen Z, et al. Efficacy and safety of sorafenib in patients in the Asia-Pacific region with advanced hepatocellular carcinoma: a phase III randomised, double-blind, placebo-controlled trial. Lancet Oncol 2009;10:25-34. Crossref

11. Llovet JM, Ricci S, Mazzaferro V, et al. Sorafenib in advanced hepatocellular carcinoma. N Engl J Med 2008;359:378-90. Crossref

12. Iavarone M, Cabibbo G, Piscaglia F, et al. Field-practice study of sorafenib therapy for hepatocellular carcinoma: a prospective multicenter study in Italy. Hepatology 2011;54:2055-63. Crossref

13. Marrero JA, Kudo M, Venook AP, et al. Observational registry of sorafenib use in clinical practice across Child-Pugh subgroups: the GIDEON study. J Hepatol 2016;65:1140-7. Crossref

14. Hiraoka A, Michitaka K, Kumada T, et al. Validation and potential of albumin–bilirubin grade and prognostication in a nationwide survey of 46,681 hepatocellular carcinoma patients in Japan: the need for a more detailed evaluation of hepatic function. Liver Cancer 2017;6:325-36. Crossref

15. Yu SC, Hui JW, Hui EP, et al. Unresectable hepatocellular carcinoma: randomized controlled trial of transarterial ethanol ablation versus transcatheter arterial chemoembolization. Radiology 2014;270:607-20. Crossref

16. Yu SC, Hui JW, Li L, et al. Comparison of chemoembolization, radioembolization, and transarterial ethanol ablation for huge hepatocellular carcinoma (≥10 cm) in tumour response and long-term survival outcome. Cardiovasc Intervent Radiol 2022;45:172-81. Crossref

17. Cheng AL, Qin S, Ikeda M, et al. Updated efficacy and safety data from IMbrave150: atezolizumab plus bevacizumab vs. sorafenib for unresectable hepatocellular carcinoma. J Hepatol 2022;76:862-73. Crossref

18. Ren Z, Xu J, Bai Y, et al. Sintilimab plus a bevacizumab biosimilar (IBI305) versus sorafenib in unresectable hepatocellular carcinoma (ORIENT-32): a randomised, open-label, phase 2-3 study. Lancet Oncol 2021;22:977-90. Crossref

19. Abou-Alfa GK, Chan SL, Kudo M, et al. Phase 3 randomized, open-label, multicenter study of tremelimumab (T) and durvalumab (D) as first-line therapy in patients (pts) with unresectable hepatocellular carcinoma (uHCC): HIMALAYA. J Clin Oncol 2022;40(4_suppl):379. Crossref

20. Yau T, Park JW, Finn RS, et al. Nivolumab versus sorafenib in advanced hepatocellular carcinoma (CheckMate 459): a randomised, multicentre, open-label, phase 3 trial. Lancet Oncol 2022;23:77-90. Crossref

21. Kudo M, Finn RS, Qin S, et al. Lenvatinib versus sorafenib in first-line treatment of patients with unresectable hepatocellular carcinoma: a randomised phase 3 non-inferiority trial. Lancet 2018;391:1163-73. Crossref

22. Ogasawara S, Ooka Y, Koroki K, et al. Switching to systemic therapy after locoregional treatment failure: definition and best timing. Clin Mol Hepatol 2020;26:155-62. Crossref

23. Kudo M. A novel treatment strategy for patients with intermediate-stage HCC who are not suitable for TACE: upfront systemic therapy followed by curative conversion. Liver Cancer 2021;10:539-44. Crossref

24. Kudo M. Extremely high objective response rate of lenvatinib: its clinical relevance and changing the treatment paradigm in hepatocellular carcinoma. Liver Cancer 2018;7:215-24. Crossref

25. Kudo M, Han KH, Ye SL, et al. A changing paradigm for the treatment of intermediate-stage hepatocellular carcinoma: Asia-Pacific Primary Liver Cancer Expert Consensus Statements. Liver Cancer 2020;9:245-60. Crossref

26. Kudo M, Kawamura Y, Hasegawa K, et al. Management of hepatocellular carcinoma in Japan: JSH Consensus Statements and Recommendations 2021 update. Liver Cancer 2021;10:181-223. Crossref

27. Hung YW, Lee IC, Chi CT, et al. Redefining tumor burden in patients with intermediate-stage hepatocellular carcinoma: the seven-eleven criteria. Liver Cancer 2021;10:629-40. Crossref

28. Kim JH, Shim JH, Lee HC, et al. New intermediate-stage subclassification for patients with hepatocellular carcinoma treated with transarterial chemoembolization. Liver Int 2017;37:1861-8. Crossref

29. Lee IC, Hung YW, Liu CA, et al. A new ALBI-based model to predict survival after transarterial chemoembolization for BCLC stage B hepatocellular carcinoma. Liver Int 2019;39:1704-12. Crossref

Peripartum cardiomyopathy (PPCM) is a rare form of heart failure that occurs in relation to pregnancy, resulting in substantial morbidity and mortality.1 In 2010, the Heart Failure Association of the European Society of Cardiology (ESC) defined PPCM as “an idiopathic cardiomyopathy presenting with heart failure secondary to left ventricular systolic dysfunction towards the end of pregnancy or in the months following delivery, where no other cause of heart failure is found”.2 Globally, its incidence varies widely, ranging from 1 in 100 live births in Nigeria3 to 1 in 20 000 live births in Japan.4

The exact pathogenesis of PPCM is not yet fully understood; the current hypothesis proposes a ‘two-hit’ model involving an initial vascular insult caused by vasculotoxic hormonal effects, including soluble FMS-like tyrosine kinase-1 and prolactin, followed by a second hit of underlying predisposition—such as genetic susceptibility and other risk factors—that limits some women’s ability to withstand this vasculotoxic insult.1 Genetic or familial predisposition to PPCM has been supported by multiple reports.5 6 7 8 Additionally, well-recognised risk factors for PPCM include advanced maternal age, African American ancestry, multiple pregnancies, hypertension, and pre-eclampsia.9

Peripartum cardiomyopathy is a potentially life-threatening myocardial disease that affects women of all ethnic groups10 and can have long-term health consequences.11 Until now, there has been a lack of information regarding the clinical phenotype and outcomes of this disease in Hong Kong. The present population-based study was conducted to evaluate the local incidence, clinical presentation, management, complications, 12-month outcomes, and subsequent pregnancies in women with PPCM. Additionally, we examined potential risk factors by comparing the clinical characteristics of women with and without PPCM to provide a basis for future preventive strategies.

This was a population-based retrospective study of all women with PPCM who delivered in public hospitals in Hong Kong between 1 January 2013 and 31 December 2022. Cases were identified through the Clinical Data Analysis and Reporting System, which captures obstetric data and hospitalisation diagnoses from eight public hospitals providing obstetric services. First, all women who delivered during the study period and had a diagnosis code for heart failure from the third trimester to 6 months postpartum were identified. Each woman’s medical record was systematically reviewed by two authors to determine whether the following criteria for PPCM were met: development of cardiac failure (with left ventricular ejection fraction [LVEF] <45% on echocardiography) during the third trimester or within 6 months postpartum without an identifiable cause. Women were excluded if LVEF was ≥45%, a recognised cause of heart failure was identified, or there was no physician-confirmed diagnosis of PPCM.

Baseline characteristics (including socio-demographics, preexisting health conditions, and obstetric history) at the time of PPCM diagnosis were obtained from medical records. Clinical presentation and initial investigations, including electrocardiography, chest radiography, echocardiography, and laboratory results, were collected. All in-hospital complications and reported outcomes during follow-up were recorded, including all-cause mortality and cardiac recovery determined by echocardiography at 12 months. Management strategies were documented, including admission to the intensive care unit or cardiac care unit, use of mechanical ventilation or circulatory support, medications prescribed at hospital discharge, pacemaker insertion, and heart transplantation. Complete recovery of cardiac function was defined as LVEF ≥50%. Some patients underwent genetic evaluation, and their reports were analysed.

Obstetric outcomes at the time of the PPCM event were assessed, including hypertensive disorders of pregnancy; gestational diabetes; thyroid disease; antenatal anaemia (defined as a haemoglobin level <10.5 g/dL); use of tocolytics; placenta accreta spectrum; placental abruption; fetal growth restriction; preterm delivery; assisted vaginal delivery or caesarean section; primary postpartum haemorrhage (blood loss ≥500 mL); and caesarean hysterectomy. Neonatal outcomes were examined, including stillbirth, sex, birth weight, small for gestational age, Apgar scores, admission to the neonatal intensive care unit, and death within 28 days of life. Data from the territory-wide electronic healthcare database were also extracted regarding outcomes of subsequent pregnancies, including LVEF before, during, and after pregnancy. The interval between the PPCM pregnancy and the first subsequent pregnancy was recorded.

To investigate risk factors for PPCM, women who gave birth during the same period but did not develop heart failure were selected as the control group, with a PPCM-to-control ratio of 1:4. Demographic and clinical characteristics were compared between women with and without PPCM.

Data analysis was conducted using SPSS (Windows version 26.0; IBM Corp, Armonk [NY], United States). The incidence rate was calculated by dividing the total number of PPCM cases by the total number of live births during the study period. Descriptive data for continuous variables were presented as mean ± standard deviation or median (range or interquartile range), and categorical data were presented as numbers with percentages. Comparisons between women with and without PPCM were performed using Student’s t test or the Mann-Whitney U test for continuous variables, and the Chi squared test or Fisher’s exact test for categorical variables. Risk factors associated with PPCM were assessed using univariable and multivariable logistic regression analyses, with results expressed as odds ratios (ORs) and 95% confidence intervals (95% CIs). A P value of <0.05 was considered statistically significant. The STROBE (Strengthening the Reporting of Observational Studies in Epidemiology) guidelines were followed in the preparation of this article.

During the 10-year study period, 30 women with PPCM delivered in public hospitals (Fig 1). Over the same period, there were 335 376 live births, yielding an estimated PPCM incidence of 1 in 11 179 live births in Hong Kong.

Detailed characteristics are listed in Table 1. All women in this study were Asian. The mean age was 33.5 years and the median body mass index was 22.0 kg/m². One woman had a positive family history of heart failure of unknown cause; no women had a previous history of PPCM or cardiac disease.

Symptoms began antepartum in 36.7% of women and postpartum in 63.3%; PPCM was predominantly diagnosed postpartum (83.3%). The median time from symptom onset to diagnosis was 3.5 days (range, 0-107). At diagnosis, 90% of women had severe symptoms (New York Heart Association functional class III/IV), most commonly comprising shortness of breath, peripheral oedema, and desaturation. Common electrocardiographic findings included sinus tachycardia and prolonged QTc interval. At the first echocardiographic assessment, the median LVEF was 30% (range, 10-44). More than half of the women had abnormal chest radiographs showing congestive lung fields, cardiomegaly, and pleural effusion (Table 2).

Detailed results are presented in Table 3. Of the 30 women with PPCM, 19 (63.3%) were managed in the intensive care unit or cardiac care unit. Cardiogenic shock, respiratory failure, and acute renal failure occurred in 10% to 20% of cases. Inotropic support, mechanical ventilation, extracorporeal membrane oxygenation, and renal replacement therapy were used during acute treatment.

At hospital discharge, most women were prescribed angiotensin-converting enzyme inhibitors (ACEis) or angiotensin receptor blockers (ARBs) and beta-blockers. Four women received prophylactic low–molecular-weight heparin for venous thromboembolism prevention after the event; another four required warfarin for the treatment of cerebral venous thrombosis, brachial artery thromboembolism, pulmonary embolism, or deep vein thrombosis (Table 3).

One woman experienced decompensated heart failure requiring an intra-aortic balloon pump and a left ventricular assist device 9 months after diagnosis, followed by heart transplantation 1 year after the event. Two women underwent implantable cardioverter-defibrillator insertion due to symptomatic premature ventricular contractions and poor LVEF recovery. Seven women (23.3%) experienced nine thromboembolic events within 1 year of the PPCM episode, including left ventricular thrombi, ischaemic stroke, and pulmonary embolism. The median follow-up duration after PPCM was 47 months (range, 3-140). At 12 months, all-cause in-hospital mortality was 6.7%; causes of death were myocardial infarction and pulmonary embolism. Overall, recovery of left ventricular function (LVEF ≥50%) occurred in 60% of women (Table 3).

Prior to PPCM, 80% of women received antenatal care. Four women (13.3%) had twin pregnancies. Antenatal anaemia was present in 50% of women. Hypertensive disorders of pregnancy occurred in 56.7%, whereas gestational diabetes was noted in 13.3%. Complications related to pre-eclampsia included haemolysis, elevated liver enzymes, and low platelets syndrome in 3.3%; eclampsia in 3.3%; and placental abruption in 6.7%. No women received tocolytics during pregnancy. The median gestational age at delivery was 37 weeks (range, 28-41). The caesarean section rate was 53.3%, and the most frequent indication was unstable maternal condition (31.3%). Primary postpartum haemorrhage occurred in 30% of cases; one woman required hysterectomy for placenta accreta spectrum. Among the 34 newborns, 32 (94.1%) were born alive; two were stillborn in the third trimester (5.9%) due to placental abruption and trisomy 18. The median birth weight was 2745 g, and 11.8% of newborns were small for gestational age. Four newborns (11.8%) had an Apgar score below 7 at 5 minutes, and nine (26.5%) required admission to a neonatal intensive care unit. There were no cases of early neonatal death (Table 4).

The obstetric and cardiac outcomes of the 11 women with subsequent pregnancies are shown in Figure 2. The median interval between the PPCM-affected pregnancy and the next pregnancy was 17 months (range, 4-60). There were 13 subsequent pregnancies (three miscarriages, five terminations, and five live births). Of the five terminations, two were advised due to poor cardiac condition; the remaining three were elective for maternal anxiety or social reasons. There were no maternal deaths or cases of recurrent PPCM.

Genetic analysis using a dilated cardiomyopathy (DCM) panel by next-generation sequencing was requested by physicians in three cases (online supplementary Table 1). Case 1, involving a woman with a family history of heart failure, revealed a pathogenic variant in the FLNC gene. Case 2, concerning a patient with a history of cancer-related chemotherapy who developed refractory postpartum heart failure requiring heart transplantation 1 year after PPCM diagnosis, had no prior signs of heart failure before pregnancy. A genetic test identified two pathogenic variants in the TTN and MYBPC3 genes. Case 3 involved a woman with chronic kidney disease who exhibited persistent left ventricular systolic dysfunction 4 years after PPCM diagnosis. Genetic evaluation was pursued due to her young-onset multisystem disease, revealing a variant in the NEXN gene. This variant, associated with autosomal dominant monogenic DCM, was absent from population databases but showed conflicting results on in silico prediction algorithms; therefore, it was classified as a variant of uncertain significance. Overall, potentially pathogenic genetic variants were identified in at least 10% of women with PPCM.

Compared with the control group, univariable logistic regression analysis showed that factors associated with PPCM included advanced maternal age (≥40 years), smoking, hypertensive disorders of pregnancy, and antenatal anaemia. In multivariable regression analysis, PPCM was independently associated with hypertensive disorders of pregnancy (adjusted OR=38.00; 95% CI=9.66-149.52; P<0.001) and antenatal anaemia (adjusted OR=13.04; 95% CI=3.72-45.70; P<0.001) [online supplementary Table 2].

Over the 10-year study period, we observed a PPCM incidence of 1 in 11 179 live births in Hong Kong. Worldwide variation in PPCM incidence may relate to ethnic and socio-economic factors12; rates are expected to increase because of advancing maternal age,13 multiple pregnancies, and obesity. About one-third of our patients developed symptoms before delivery, a finding comparable to the Asia-Pacific group in the ESC EURObservational Research Programme registry.10 Overall, 30% of women were diagnosed more than 7 days after symptom onset. Among those with antepartum-onset symptoms, 54.5% were diagnosed after delivery. This diagnostic delay may be attributed to the difficulty in distinguishing PPCM from normal physiological changes of pregnancy—its symptoms often mimic those of late gestation and may only be recognised postpartum when they become more pronounced. Delayed diagnosis has been associated with lower rates of left ventricular recovery.14 Early recognition and awareness among both pregnant women and healthcare professionals are crucial to enable prompt initiation of heart failure therapy, which may improve cardiac recovery. To support early detection and facilitate timely specialist referral for diagnostic evaluation, serum biomarkers can be measured to rule out heart failure with high probability during pregnancy or the postpartum period.15

In our study, approximately half of the cases involved pre-eclampsia, a finding consistent with the Asia-Pacific cohort in the ESC EURObservational Research Programme registry.10 A meta-analysis of 22 studies demonstrated a fourfold higher prevalence of pre-eclampsia among women with PPCM relative to the general obstetric population (22% vs 5%).16 Our multivariable regression analysis confirmed that hypertensive disorders of pregnancy constituted an independent risk factor for PPCM. The association between pre-eclampsia and PPCM may be explained by their shared pathophysiological mechanism—systemic vascular angiogenic imbalance.1 15 17 Preeclampsia and PPCM might represent a single disease spectrum with substantial overlap.17 Low-dose aspirin is generally used for the prevention of pre-eclampsia and its associated morbidity and mortality.18 Although aspirin use for PPCM prevention is not supported by evidence-based guidelines, it could theoretically provide benefit due to the shared vascular dysfunction pathways. Consequently, the use of aspirin for pre-eclampsia prevention may indirectly reduce the risk of PPCM in high-risk women.

We found that antenatal anaemia was independently associated with PPCM. A systematic review and meta-analysis previously indicated that women with anaemia had up to fivefold higher odds of developing PPCM compared with women exhibiting normal haemoglobin levels.19 The precise nature of this association remains unclear; iron deficiency may contribute by impairing myocardial contractile function.20 Anaemia screening and correction during pregnancy may help reduce the risk of PPCM.

A multidisciplinary approach involving cardiologists, obstetricians, intensivists, cardiac surgeons, anaesthesiologists, neonatologists, and nurses is essential for the management of PPCM.21 In severe cases with haemodynamic instability, acute management—including immediate resuscitation and mechanical respiratory or circulatory support—may be required.15 Urgent caesarean section should be considered for advanced heart failure that persists despite optimal medical therapy. According to international consensus, the main treatment should follow guideline-directed medical therapy for heart failure with reduced ejection fraction in non-pregnant patients, while respecting contraindications for certain drugs during pregnancy.6 22 23 24 25 Standard therapies include diuretics, ACEis or ARBs, mineralocorticoid receptor antagonists, vasodilators (hydralazine/nitrates), digoxin, beta-blockers, and anticoagulants. A 2022 meta-analysis of global data demonstrated that frequent prescription of beta-blockers, ACEis/ARBs, and bromocriptine or cabergoline was associated with lower all-cause mortality and better left ventricular recovery at 12 months.26 In our study, most patients received ACEis/ARBs and beta-blockers; fewer were prescribed bromocriptine at discharge. The rationale for using dopamine agonists to inhibit prolactin secretion lies in the proposed pathophysiological mechanism involving 16-kDa prolactin, an oxidative stress-mediated cleavage product that damages cardiovascular tissue.27 Regarding prolactin inhibition in women with PPCM, a meta-analysis reported that those treated with bromocriptine had higher odds of left ventricular recovery, without a significant difference in all-cause mortality.28 However, bromocriptine use is associated with an increased risk of thromboembolic complications. The 2019 ESC–Heart Failure Association position statement issued a weak recommendation for bromocriptine use, advising that it should always be accompanied by at least prophylactic anticoagulation.15 Future randomised controlled trials and registry data with longer follow-up are needed to provide stronger evidence supporting its use. For women who do not recover from PPCM within 1 year, the American College of Cardiology/American Heart Association Joint Committee and the ESC recommend implantable cardioverter-defibrillator therapy for the primary prevention of sudden cardiac death due to ventricular tachyarrhythmia.22 29 30 Cardiac transplantation may be required for patients with refractory severe heart failure despite maximal medical therapy, as occurred in one of our cases.

Estimates of left ventricular recovery and mortality in PPCM vary considerably across geographic regions,26 presumably due to differences in medical therapy, access to healthcare services, and follow-up duration. A 2022 meta-analysis of 4875 patients from 60 countries reported overall 12-month rates of left ventricular recovery and all-cause mortality of 58.7% and 9.8%, respectively.26 In our cohort, 60% of women achieved cardiac recovery; two patients (6.7%) died of myocardial infarction and pulmonary embolism within 12 months of diagnosis. Both had poor social support and did not adhere to treatment or attend follow-up visits, which likely contributed to their adverse outcomes. These findings highlight the need for greater public awareness, improved medication compliance, and stronger social support systems. We recommend enhanced nursing outreach and structured patient education, along with post-discharge monitoring, to optimise outcomes.

Thromboembolism, a potentially life-threatening complication of PPCM, affected 23.3% of women in our cohort. This high rate may be attributed to the hypercoagulable state of pregnancy, impaired circulation, and blood stasis from cardiac failure. Our incidence was higher than the reported global rate of 6.1% in a recent international study.26 Therapeutic anticoagulation is recommended for patients with intracardiac thrombus or systemic embolism. In our study, 13.3% of patients received low molecular weight heparin for thromboembolism prophylaxis. Both the AHA and ESC recommend anticoagulation in PPCM cases involving severe left ventricular dysfunction (LVEF <30% to <35%) during the peripartum period and up to 8 weeks postpartum.29 31 Despite the high thromboembolic risk in PPCM, anticoagulation remains a subject of ongoing debate.32 Our data support prophylactic anticoagulation for all women with PPCM, given the high incidence observed. Ultimately, individual assessment of thromboembolic risk—considering the extent of left ventricular dysfunction, caesarean delivery, immobility, and ventricular dilatation—may help identify patients most likely to benefit from thromboprophylaxis.

Relapse of PPCM and associated mortality in subsequent pregnancies are not uncommon; rates range from 5.3% to 29.5% and 0% to 55.5%, respectively.33 In our study, nine of 11 patients (81.8%) had confirmed recovery of cardiac function before conception. There were no maternal deaths or PPCM recurrences during pregnancy. A recent meta-analysis showed that women with persistent left ventricular dysfunction prior to a subsequent pregnancy had a higher risk of mortality and worsening function compared to women whose cardiac function had recovered.33 However, recovered left ventricular function does not guarantee an uncomplicated subsequent pregnancy.34 35 It is crucial to monitor cardiac function throughout pregnancy—and up to 6 months postpartum—to detect subclinical left ventricular dysfunction or PPCM recurrence. Women with a history of PPCM should be counselled regarding the risks of future pregnancies, including irreversible ventricular deterioration, maternal death, and fetal loss.36 Subsequent pregnancy is not recommended if LVEF fails to normalise. Contraceptive counselling should begin early after the acute event; reliable methods with minimal thromboembolic risk are preferred.37

A study has demonstrated a genetic contribution to PPCM in at least 15% of cases.38 The most commonly affected gene is TTN, which encodes the large sarcomeric protein titin.39 The relative prevalence of truncating variants in these genes is nearly identical between PPCM and DCM.39 In our study, three of 30 patients (10%) were screened for cardiomyopathy-related genes (TTN, FLNC, MYBPC3, NEXN), all of whom were in the non-recovery group, indicating that at least 10% had a genetic predisposition to PPCM. The American College of Cardiology/American Heart Association Joint Committee recommends that patients with non-ischaemic cardiomyopathy undergo genetic counselling and testing for inherited cardiomyopathies to facilitate early cardiac disease detection and timely initiation of treatments that reduce heart failure progression and sudden death risk.22 The identification of pathogenic genetic variants can provide valuable prognostic information and clarify associated risks (eg, arrhythmic complications linked to FLNC and DSP mutations), thereby guiding decisions on preventive measures, including implantable defibrillator placement and exercise recommendations. Furthermore, cascade genetic testing for relatives enables closer pregnancy monitoring, informed reproductive decisions (including prenatal or preimplantation genetic diagnosis), and lifelong cardiovascular surveillance to improve outcomes.40 The value of routine genetic testing remains limited by low penetrance, variable clinical expression, and uncertain variant significance. It may also lead to patient anxiety, potential genetic discrimination, and substantial resource implications. Careful patient selection with thorough pre- and post-test counselling is essential. Because the clinical presentation of PPCM closely resembles that of DCM, the ESC suggests that genetic testing be considered in PPCM cases with a positive family history,15 where clinically actionable findings are most likely to be identified.

This study had several limitations. Because PPCM is a rare condition, a small sample size was inevitable. The retrospective nature of data collection over a 10-year period may have resulted in incomplete information. Outcomes could also have been influenced by variations in heart failure management over time and across hospitals. Furthermore, some PPCM cases managed in the private sector or outside Hong Kong might not have been captured. The long-term impact of PPCM on women’s overall health was not assessed. The establishment of a local PPCM registry would facilitate a better understanding of the condition, identification of outcome determinants, and optimisation of clinical care in Hong Kong.

Peripartum cardiomyopathy is an uncommon but potentially life-threatening medical condition affecting women worldwide. Genetic factors contribute to disease susceptibility in at least 10% of cases. Genetic testing may offer a valuable tool to guide prognosis and management in affected women.

Concept or design: LSK Law, LT Kwong, PL So.
Acquisition of data: LSK Law, KH Siong, HC Mok, STK Wong, JKO Wai, CY Chow, WL Chan, KY Tse, YYY Chan, KS Eu, PL So.
Analysis or interpretation of data: LSK Law, PL So.
Drafting of the manuscript: LSK Law, PL So.
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 all staff in the Statistics Department at Tuen Mun Hospital for their assistance with data collection.

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 Central Institutional Review Board of Hospital Authority, Hong Kong (Ref No.: CIRB-2023-114-3). The requirement for informed patient consent was waived by the Board due to the retrospective nature of the research. All data used in the research were anonymised and unidentifiable.

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

1. Davis MB, Arany Z, McNamara DM, Goland S, Elkayam U. Peripartum cardiomyopathy: JACC state-of-the-art review. J Am Coll Cardiol 2020;75:207-21. Crossref

2. Sliwa K, Hilfiker-Kleiner D, Petrie MC, et al. Current state of knowledge on aetiology, diagnosis, management, and therapy of peripartum cardiomyopathy: a position statement from the Heart Failure Association of the European Society of Cardiology Working Group on peripartum cardiomyopathy. Eur J Heart Fail 2010;12:767-78. Crossref

3. Isezuo SA, Abubakar SA. Epidemiologic profile of peripartum cardiomyopathy in a tertiary care hospital. Ethn Dis 2007;17:228-33.

4. Kamiya CA, Kitakaze M, Ishibashi-Ueda H, et al. Different characteristics of peripartum cardiomyopathy between patients complicated with and without hypertensive disorders. -Results from the Japanese Nationwide survey of peripartum cardiomyopathy-. Circ J 2011;75:1975-81. Crossref

5. Pierce JA, Price BO, Joyce JW. Familial occurrence of postpartal heart failure. Arch Intern Med 1963;111:651-5. Crossref

6. Morales A, Painter T, Li R, et al. Rare variant mutations in pregnancy-associated or peripartum cardiomyopathy. Circulation 2010;121:2176-82. Crossref

7. van Spaendonck-Zwarts KY, van Tintelen JP, van Veldhuisen DJ, et al. Peripartum cardiomyopathy as a part of familial dilated cardiomyopathy. Circulation 2010;121:2169-75. Crossref

8. van Spaendonck-Zwarts KY, Posafalvi A, van den Berg MP, et al. Titin gene mutations are common in families with both peripartum cardiomyopathy and dilated cardiomyopathy. Eur Heart J 2014;35:2165-73. Crossref

9. Honigberg MC, Givertz MM. Peripartum cardiomyopathy. BMJ 2019;364:k5287. Crossref

10. Sliwa K, Petrie MC, van der Meer P, et al. Clinical presentation, management, and 6-month outcomes in women with peripartum cardiomyopathy: an ESC EORP registry. Eur Heart J 2020;41:3787-97. Crossref

11. Koerber D, Khan S, Kirubarajan A, et al. Meta-analysis of long-term (>1 year) cardiac outcomes of peripartum cardiomyopathy. Am J Cardiol 2023;194:71-7. Crossref

12. Karaye KM, Ishaq NA, Sai’du H, et al. Disparities in clinical features and outcomes of peripartum cardiomyopathy in high versus low prevalent regions in Nigeria. ESC Heart Fail 2021;8:3257-67. Crossref

13. Kolte D, Khera S, Aronow WS, et al. Temporal trends in incidence and outcomes of peripartum cardiomyopathy in the United States: a nationwide population-based study. J Am Heart Assoc 2014;3:e001056. Crossref

14. Lewey J, Levine LD, Elovitz MA, Irizarry OC, Arany Z. Importance of early diagnosis in peripartum cardiomyopathy. Hypertension 2020;75:91-7. Crossref

15. Bauersachs J, König T, van der Meer P, et al. Pathophysiology, diagnosis and management of peripartum cardiomyopathy: a position statement from the Heart Failure Association of the European Society of Cardiology Study Group on peripartum cardiomyopathy. Eur J Heart Fail 2019;21:827-43. Crossref

16. Bello N, Rendon IS, Arany Z. The relationship between pre-eclampsia and peripartum cardiomyopathy: a systematic review and meta-analysis. J Am Coll Cardiol 2013;62:1715-23. Crossref

17. Parikh P, Blauwet L. Peripartum cardiomyopathy and preeclampsia: overlapping diseases of pregnancy. Curr Hypertens Rep 2018;20:69. Crossref

18. Henderson JT, Vesco KK, Senger CA, Thomas RG, Redmond N. Aspirin use to prevent preeclampsia and related morbidity and mortality: updated evidence report and systematic review for the US Preventive Services Task Force. JAMA 2021;326:1192-206. Crossref

19. Cherubin S, Peoples T, Gillard J, Lakhal-Littleton S, Kurinczuk JJ, Nair M. Systematic review and meta-analysis of prolactin and iron deficiency in peripartum cardiomyopathy. Open Heart 2020;7:e001430. Crossref

20. Anand IS, Gupta P. Anemia and iron deficiency in heart failure: current concepts and emerging therapies. Circulation 2018;138:80-98. Crossref

21. Sigauke FR, Ntsinjana H, Tsabedze N. Peripartum cardiomyopathy: a comprehensive and contemporary review. Heart Fail Rev 2024;29:1261-78. Crossref

22. Heidenreich PA, Bozkurt B, Aguilar D, et al. 2022 AHA/ACC/HFSA Guideline for the Management of Heart Failure: a report of the American College of Cardiology/American Heart Association Joint Committee on Clinical Practice Guidelines. Circulation 2022;145:e895-1032. Crossref

23. Arany Z. Peripartum cardiomyopathy. N Engl J Med 2024;390:154-64. Crossref

24. Azibani F, Sliwa K. Peripartum cardiomyopathy: an update. Curr Heart Fail Rep 2018;15:297-306. Crossref

25. Maddox TM, Januzzi JL Jr, Allen LA, et al. 2024 ACC Expert Consensus Decision Pathway for treatment of heart failure with reduced ejection fraction: a report of the American College of Cardiology Solution Set Oversight Committee. J Am Coll Cardiol 2024;83:1444-88. Crossref

26. Hoevelmann J, Engel ME, Muller E, et al. A global perspective on the management and outcomes of peripartum cardiomyopathy: a systematic review and meta-analysis. Eur J Heart Fail 2022;24:1719-36. Crossref

27. Hilfiker-Kleiner D, Kaminski K, Podewski E, et al. A cathepsin D–cleaved 16 kDa form of prolactin mediates postpartum cardiomyopathy. Cell 2007;128:589-600. Crossref

28. Kumar A, Ravi R, Sivakumar RK, et al. Prolactin inhibition in peripartum cardiomyopathy: systematic review and meta-analysis. Curr Probl Cardiol 2023;48:101461. Crossref

29. Bauersachs J, Arrigo M, Hilfiker-Kleiner D, et al. Current management of patients with severe acute peripartum cardiomyopathy: practical guidance from the Heart Failure Association of the European Society of Cardiology Study Group on peripartum cardiomyopathy. Eur J Heart Fail 2016;18:1096-105. Crossref

30. McDonagh TA, Metra M, Adamo M, et al. 2021 ESC Guidelines for the diagnosis and treatment of acute and chronic heart failure: developed by the Task Force for the diagnosis and treatment of acute and chronic heart failure of the European Society of Cardiology (ESC). With the special contribution of the Heart Failure Association (HFA) of the ESC. Eur J Heart Fail 2022;24:4-131. Crossref

31. Bozkurt B, Colvin M, Cook J, et al. Current diagnostic and treatment strategies for specific dilated cardiomyopathies: a scientific statement from the American Heart Association. Circulation 2016;134:e579-646. Crossref

32. Radakrishnan A, Dokko J, Pastena P, Kalogeropoulos AP. Thromboembolism in peripartum cardiomyopathy: a systematic review. J Thorac Dis 2024;16:645-60. Crossref

33. Wijayanto MA, Myrtha R, Lukas GA, et al. Outcomes of subsequent pregnancy in women with peripartum cardiomyopathy: a systematic review and meta-analysis. Open Heart 2024;11:e002626. Crossref

34. Pachariyanon P, Bogabathina H, Jaisingh K, Modi M, Modi K. Long-term outcomes of women with peripartum cardiomyopathy having subsequent pregnancies. J Am Coll Cardiol 2023;82:16-26. Crossref

35. Fett JD, Shah TP, McNamara DM. Why do some recovered peripartum cardiomyopathy mothers experience heart failure with a subsequent pregnancy? Curr Treat Options Cardiovasc Med 2015;17:354. Crossref

36. Sliwa K, van der Meer P, Petrie MC, et al. Corrigendum to ‘Risk stratification and management of women with cardiomyopathy/heart failure planning pregnancy or presenting during/after pregnancy: a position statement from the Heart Failure Association of the European Society of Cardiology Study Group on Peripartum Cardiomyopathy’ [Eur J Heart Fail 2021;23:527-540]. Eur J Heart Fail 2022;24:733. Crossref

37. Sliwa K, Petrie MC, Hilfiker-Kleiner D, et al. Long-term prognosis, subsequent pregnancy, contraception and overall management of peripartum cardiomyopathy: practical guidance paper from the Heart Failure Association of the European Society of Cardiology Study Group on Peripartum Cardiomyopathy. Eur J Heart Fail 2018;20:951-62. Crossref

38. Ware JS, Li J, Mazaika E, et al. Shared genetic predisposition in peripartum and dilated cardiomyopathies. N Engl J Med 2016;374:233-41. Crossref

39. Goli R, Li J, Brandimarto J, et al. Genetic and phenotypic landscape of peripartum cardiomyopathy. Circulation 2021;143:1852-62. Crossref

40. Arany Z. It is time to offer genetic testing to women with peripartum cardiomyopathy. Circulation 2022;146:4-5. Crossref

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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