The Global Burden of Disease Results Tool of the Institute for Health Metrics and Evaluation reported that cardiovascular diseases accounted for approximately 10 383 550 deaths globally in 2017, representing 18.56% of all-cause mortality.1 Cardiothoracic surgery plays an important role in the treatment of these conditions and in reducing associated morbidity and mortality. However, surgery carries inherent risks that vary among patients, necessitating careful evaluation of risks and benefits before proceeding. A risk stratification tool is essential for effective patient triage and the consent process.

One widely used risk stratification tool is the European System for Cardiac Operative Risk Evaluation (EuroSCORE), a specialised scoring system that provides customised predictions of in-hospital mortality after cardiac surgery. The tool assigns scores based on various preoperative risk factors to stratify patients into different risk categories (low: EuroSCORE <4%, intermediate: 4-8%, high: >8%).2 In the UK, in-hospital mortality declined from 4.0% to 2.8% between 2002 and 2016 following implementation of EuroSCORE,3 supporting its value in cardiac surgical risk assessment. The EuroSCORE comprises three versions: the additive EuroSCORE,4 the logistic EuroSCORE,5 and EuroSCORE II.6 In 2012, the Society for Cardiothoracic Surgery in Great Britain and Ireland recommended the use of the latest version, EuroSCORE II.6

Prince of Wales Hospital (PWH) in Hong Kong has adopted the logistic EuroSCORE for risk assessment since 2007. However, several publications from different countries have raised concerns regarding the accuracy of the additive and logistic EuroSCORE models, leading to the development of EuroSCORE II.7 8 9 Consequently, EuroSCORE II has been proposed as the future risk adjustment tool of the Society for Cardiothoracic Surgery in Great Britain and Ireland following successful contemporary validation.6 10 11

Although EuroSCORE II has been widely used and validated, the underlying data were predominantly derived from Western populations undergoing cardiac surgery in Europe and the US.4 12 13 14 Therefore, studies evaluating the performance of EuroSCORE II in Asian populations remain limited,15 16 17 and none has been conducted specifically in Hong Kong. Furthermore, no studies have compared the performance of the logistic EuroSCORE and EuroSCORE II in the Hong Kong population. The present study aimed to address these gaps.

Moreover, Hong Kong has a higher proportion of aortic surgery than Western countries. Prince of Wales Hospital reported a surge in aortic surgeries, reaching 26% between 2021 and 2022,18 whereas the UK reported an aortic surgery prevalence of 3.47% between 2015 and 2016.3 We therefore sought to investigate whether this variation influences the validity of EuroSCORE II through subgroup analyses.

The primary objective of this study was to assess the discriminatory ability and calibration of EuroSCORE II in predicting postoperative mortality after the three main index cardiac surgeries (ie, coronary artery bypass grafting [CABG], valve surgery, and aortic surgery) at our centre.2 7 8 9 14 15 16 17 19 20 21 22 23 The secondary objective was to compare the discriminatory ability and calibration of EuroSCORE II with those of the logistic EuroSCORE in patients undergoing cardiac surgery.

This retrospective validation study included patients (aged ≥18 years) who underwent all types of cardiac surgery—including CABG, valve surgery (eg, aortic valve replacement, mitral valve replacement, and tricuspid valve repair), aortic surgery, isolated or combined procedures, and other procedures (eg, left atrial appendage closure) at PWH between 1 January 2013 and 31 December 2023 (inclusive) [Fig]. Because PWH does not perform certain cardiothoracic procedures—such as paediatric cardiac surgery, cardiac transplantation, and oesophageal surgery—records for these interventions were unavailable. For patients who underwent multiple cardiac surgeries during the same hospital admission, only the first index procedure was analysed. The minimum sample size of 225 was calculated based on estimates of area under the receiver operating characteristic curve (AUROC) from the literature7 and the estimated prevalence of the outcome (online supplementary Table),7 24 indicating that our primary cohorts for CABG, valve surgery, and aortic surgery exceeded the required sample size.

The Dendrite cardiac surgery database (Dendrite Clinical Systems, Oxford, United Kingdom)25 was utilised for secondary data collection (Fig). This database captures clinically relevant information, including preoperative medical records and postoperative complications for patients undergoing cardiac surgery. All key variables required to calculate EuroSCORE II6 and the logistic EuroSCORE5 were extracted. Mortality, the primary outcome, was defined as death within 30 days of the index operation (regardless of place of death), consistent with previous studies.2 9 14 15 21

Cases with missing or incomplete data required for calculation of EuroSCORE II were excluded from the analysis. Analyses were performed for the overall cohort and stratified by individual cardiac procedure. Complete data were available for validation of EuroSCORE II and comparison of its predictive performance with that of the logistic EuroSCORE for postoperative mortality.

Univariate and multivariate binary logistic regression analyses were conducted on all relevant variables included in the EuroSCORE II scale to identify significant covariates associated with an increased risk of mortality.

The discriminatory performance of the predictive models was evaluated using the AUROC; values of 0.8 or above indicated strong discrimination, and 1.0 indicated perfect discrimination. Pairwise comparisons of AUROCs for individual cardiac procedures were performed using the DeLong test, with the threshold for statistical significance set at P<0.05.

Calibration of the predictive model was evaluated using the Hosmer–Lemeshow goodness-of-fit test and calibration plots, through statistical and graphical assessment of agreement between observed and expected event rates within model subgroups. A P value >0.05 and a regression line approximating the 45-degree diagonal indicated good calibration, reflecting adequate agreement between observed and predicted event rates.

Model goodness of fit was further assessed using the coefficient of determination (R²), which quantifies the proportion of variance explained by the model, and the normalised root mean square error (NRMSE), which measures predictive accuracy by comparing predicted and observed values, normalised to the data range. Higher R² values and lower NRMSE values indicate better model fit.

Statistical analyses were performed using SPSS (Windows version 29.0; IBM Corp, Armonk [NY], United States), Microsoft Excel 2019, and R software (RStudio, version 2024.04.2).

The study cohort comprised 4180 patients (Fig). Table 1 summarises the characteristics of the overall cohort and relevant subgroups. The median age was 63 years (interquartile range, 56-69), and 29.2% (n=1222) were women. Aortic operations were performed in 21.1% (n=883) of patients and the majority underwent a single non-coronary procedure (36.9%, n=1541). For the overall cohort, the median logistic EuroSCORE value was 5.8 (interquartile range, 2.6-13.7), whereas median EuroSCORE II value was 2.4 (interquartile range, 1.2-5.3). The institutional 30-day mortality rate for all cardiac procedures was 4.2%.

The AUROC for EuroSCORE II was 0.829, indicating strong discriminatory ability. The Hosmer–Lemeshow P value for EuroSCORE II was 0.155, indicating no statistically significant difference between predicted and observed values (online supplementary Fig a). Accordingly, EuroSCORE II demonstrated acceptable calibration.

EuroSCORE II demonstrated a statistically significant improvement in discriminatory performance compared with the logistic EuroSCORE (DeLong P=0.006). Additionally, EuroSCORE II showed superior calibration, supported by a significant Hosmer–Lemeshow test result for the logistic EuroSCORE (P<0.001). Calibration curves comparing observed and predicted 30-day mortality were consistent with these findings, further indicating better calibration with EuroSCORE II than with the logistic EuroSCORE. More than 90% of the variation in 30-day mortality was explained by both models (R² for EuroSCORE II=98.7%; R² for logistic EuroSCORE=99.1%). Notably, EuroSCORE II demonstrated a substantially lower NRMSE (5.7%) than the logistic EuroSCORE (56.4%), indicating reduced dispersion and relative variability in predictions (online supplementary Fig a).

In this subgroup, EuroSCORE II demonstrated strong discriminatory performance (AUROC=0.847) and acceptable calibration (Hosmer–Lemeshow P=0.113). There was no statistically significant difference in discriminatory performance between EuroSCORE II and the logistic EuroSCORE (DeLong P=0.529). However, EuroSCORE II showed better calibration, supported by a significant Hosmer–Lemeshow test result for the logistic EuroSCORE (P<0.001) and calibration curves favouring EuroSCORE. More than 85% of the variation in 30-day mortality was explained by both models (R2 for EuroSCORE II=87.7%; R² for logistic EuroSCORE=91.4%). Compared with the logistic EuroSCORE (40.6%), EuroSCORE II demonstrated a lower NRMSE (13.0%), indicating reduced dispersion and relative variability (online supplementary Fig b).

In this subgroup, EuroSCORE II demonstrated strong discriminatory performance (AUROC=0.810) and acceptable calibration (Hosmer–Lemeshow P=0.162). There was no statistically significant difference in discriminatory performance between EuroSCORE II and the logistic EuroSCORE (DeLong P=0.160). Nevertheless, EuroSCORE II demonstrated superior calibration, supported by a significant Hosmer–Lemeshow test result for the logistic EuroSCORE (P<0.001) and calibration curves favouring EuroSCORE II. More than 90% of the variation in 30-day mortality was explained by both models (R² for EuroSCORE II=94.7%; R² for logistic EuroSCORE=94.4%). Compared with the logistic EuroSCORE (80.4%), EuroSCORE II demonstrated a lower NRMSE (21.8%), indicating reduced dispersion and relative variability (online supplementary Fig c).

In this subgroup, EuroSCORE II demonstrated satisfactory discriminatory performance (AUROC=0.735) and good calibration (Hosmer–Lemeshow P=0.549). It also showed a statistically significant improvement in discrimination compared with the logistic EuroSCORE (DeLong P<0.001). Calibration was also superior, supported by a significant Hosmer–Lemeshow test result for the logistic EuroSCORE (P<0.001) and calibration curves favouring EuroSCORE II. More than 90% of the variation in 30-day mortality was explained by EuroSCORE II (R² for EuroSCORE II=96.1%; R² for logistic EuroSCORE=76.6%). EuroSCORE II also demonstrated a lower NRMSE (6.6%) than the logistic EuroSCORE (98.8%), indicating reduced dispersion and relative variability (online supplementary Fig d).

In this subgroup, EuroSCORE II demonstrated fair discriminatory performance (AUROC=0.694) and good calibration (Hosmer–Lemeshow P=0.606). There was no statistically significant difference in discriminatory performance between EuroSCORE II and the logistic EuroSCORE (DeLong P=0.913). Both models exhibited adequate calibration (EuroSCORE II P=0.606; logistic EuroSCORE P=0.280) [online supplementary Fig e].

In this subgroup, EuroSCORE II demonstrated strong discriminatory performance (AUROC=0.862) and acceptable calibration (Hosmer–Lemeshow P=0.159). There was no statistically significant difference in discriminatory performance between EuroSCORE II and the logistic EuroSCORE (DeLong P=0.248). However, EuroSCORE II exhibited superior calibration compared with the logistic EuroSCORE (Hosmer–Lemeshow P=0.062) [online supplementary Fig f].

In this subgroup, EuroSCORE II demonstrated strong discriminatory performance (AUROC=0.872) but poor calibration (Hosmer–Lemeshow P<0.001). There was no statistically significant difference in discriminatory performance between EuroSCORE II and the logistic EuroSCORE (DeLong P=0.626). Notably, calibration curves favoured EuroSCORE II over the logistic EuroSCORE (online supplementary Fig g).

Furthermore, comparison of EuroSCORE II variables with multivariable analyses from PWH database identified dialysis as an additional significant predictor of increased 30-day mortality (adjusted odds ratio=3.401) among patients undergoing cardiac surgery (Table 2).

In the present study, EuroSCORE II demonstrated strong discriminatory performance and good calibration in the overall cohort and three key subgroups (isolated CABG, isolated valve surgery and aortic surgery). Moreover, EuroSCORE II outperformed the logistic EuroSCORE in both discrimination and calibration across the overall cohort and these principal subgroups.

Our results are consistent with validation studies conducted in several European countries (Italy,26 Greece,27 Serbia,28 Spain,29 and Hungary30), which demonstrated strong discriminatory performance (AUROC >0.7) for EuroSCORE II.26 27 28 29 30 31 These findings reaffirm the robust predictive performance of EuroSCORE II for mortality in patients undergoing cardiac surgery.

In addition to European populations, our findings align with those of validation studies conducted in Asian cohorts.15 16 17 23 Specifically, Liu et al15 demonstrated strong discriminatory performance for EuroSCORE II, with an AUROC of 0.792 in a single-centre setting. This concordance further supports the consistency and reliability of EuroSCORE II as a mortality prediction tool in Asian cardiac surgery populations.

However, Kurniawaty et al19 reported considerably different findings, demonstrating only fair discriminatory performance, with evidence of miscalibration and underprediction in an Indonesian population. This discrepancy may be attributable to differences in patient age. Both our cohort and the European cohorts had substantially higher median (63 years) or mean (64.6 years)6 ages compared with the mean age in the Indonesian cohort (44 years)19. Given the younger age profile and lower prevalence of risk factors included in the EuroSCORE II model among Indonesian patients, its predictive performance may be limited in that population. Accordingly, these findings may be less generalisable to the Hong Kong population.

For the overall cohort, EuroSCORE II demonstrated superior performance in both discrimination and calibration compared with the logistic EuroSCORE. This difference may reflect the tendency of the logistic EuroSCORE to overestimate mortality risk, particularly in high-risk emergency patients.10 Consequently, EuroSCORE II appears to provide more accurate risk stratification than the logistic EuroSCORE.

For isolated CABG procedures, the discriminatory performance of EuroSCORE II was strong in our study, supported by a non-significant Hosmer–Lemeshow statistic, consistent with findings from a large UK validation cohort.7 Studies conducted in Finland32 (AUROC=0.852) and China16 (AUROC=0.762) similarly demonstrated robust discriminatory performance of EuroSCORE II in predicting operative mortality among high-risk isolated CABG patients and those undergoing CABG with or without concomitant major cardiac surgery. However, a study from Singapore reported poor discrimination and calibration, particularly in moderate- and high-risk cohorts.33 Comparable findings were reported in studies from Indonesia22 and Malaysia,23 which demonstrated fair discrimination but underestimation of mortality after isolated CABG. These discrepancies suggest that additional caution may be warranted when applying EuroSCORE II in isolated CABG populations. Differences in demographic characteristics or study design may contribute to variability in model performance, warranting further investigation.

For aortic procedures, EuroSCORE II demonstrated higher AUROC values and more favourable Hosmer–Lemeshow P values than the logistic EuroSCORE. Nevertheless, caution is warranted because the model does not incorporate specific procedural variables (eg, open surgery vs minimally invasive approaches) as risk factors, which may limit precision in mortality prediction for aortic surgery.7

The adoption of contemporary machine learning and artificial intelligence techniques, rather than logistic regression, may offer more effective modelling approaches for capturing complex, non-linear interactions among established risk factors. Furthermore, incorporating the statistically significant variable identified through multivariate analysis of the PWH database, specifically dialysis, into a future EuroSCORE III model may further enhance its predictive performance.

First, the robustness of this validation study is supported by its substantial sample size (n=4180), which increases statistical power, enables detection of smaller effects, and enhances generalisability. Second, the absence of missing data strengthens measurement completeness and the credibility of the validation process, reduces information bias, and facilitates a more precise evaluation of EuroSCORE II predictive performance within this large cohort.

First, reliance on data from a single institution may introduce sampling bias. Therefore, multi-centre analyses should be conducted in future, provided sufficient resources are available. Second, the retrospective design limited the study by precluding long-term follow-up after patient discharge. Consequently, the analysis did not capture longer-term outcomes that may be influenced by baseline EuroSCORE II risk estimates. Additionally, the cohort demonstrated a skewed distribution across risk categories, with a substantial proportion (>85.2%) categorised as low or intermediate risk, thereby limiting generalisability to high-risk populations.

First, in aortic surgery, the discrepancy in EuroSCORE II performance observed between Hong Kong and the UK indicates a need for further investigation.10 A meta-analysis focusing on validation of EuroSCORE II in aortic procedures could help refine risk assessment in this subgroup. Second, although EuroSCORE II is a valuable risk stratification tool in cardiac surgery, minimally invasive cardiac procedures34 and certain established risk factors (eg, diffuse coronary artery disease and aortic calcification)15 are not included in the model. Accordingly, there may be a need for in-depth evaluation of their relevance to EuroSCORE II calculation. Third, multi-centre studies would enable validation of these findings on a broader scale. Collaboration with the other two cardiac centres in Hong Kong would enhance generalisability and support more robust conclusions.

In our cohort, EuroSCORE II demonstrated strong discriminatory performance and good calibration for predicting 30-day postoperative mortality among patients undergoing cardiac surgery. It also shows superior calibration and comparable or improved discrimination in the three principal subgroups—isolated CABG, isolated valve surgery, and aortic surgery—compared with the logistic EuroSCORE. Accordingly, EuroSCORE II represents a risk stratification tool superior to the logistic EuroSCORE and is well suited for use in Hong Kong.

Concept or design: KHL Ng, T Fujikawa, K Wang, RHL Wong.
Acquisition of data: KHL Ng, MWT Kwok, JYK Ho, SCY Chow, JWY Chan, K Lim, ATC Chang, ICH Siu, T Fujikawa, RHL Wong.
Analysis or interpretation of data: KHL Ng, T Fujikawa, K Wang, RHL Wong.
Drafting of the manuscript: KHL Ng, T Fujikawa, K Wang, RHL Wong.
Critical revision of the manuscript for important intellectual content: KHL Ng, T Fujikawa, RHL Wong.

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

All authors have disclosed no conflicts of interest.

The authors thank Dr Simon KS Yau from Department of Family Medicine of the New Territories East Cluster for his insightful contributions to data interpretation and manuscript revision.

This research was presented at The Hospital Authority Convention 2025 (26 May 2025, Hong Kong).

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

This research was approved by the Institutional Review Board of The Chinese University of Hong Kong/Hospital Authority New Territories East Cluster, Hong Kong (Ref No.: 2024.571). The requirement for informed patient consent was waived by the Board 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. Vervoort D, Swain JD, Pezzella AT, Kpodonu J. Cardiac surgery in low- and middle-income countries: a state-of-the-art review. Ann Thorac Surg 2021;111:1394-400. Crossref

2. Silverborn M, Nielsen S, Karlsson M. The performance of EuroSCORE II in CABG patients in relation to sex, age, and surgical risk: a nationwide study in 14,118 patients. J Cardiothorac Surg 2023;18:40. Crossref

3. Society for Cardiothoracic Surgery in Great Britain and Ireland. Blue Books. Available from: https://scts.org/professionals/reports/resources/. Accessed 10 Sep 2024.

4. Nashef SA, Roques F, Michel P, Gauducheau E, Lemeshow S, Salamon R. European System for Cardiac Operative Risk Evaluation (EuroSCORE). Eur J Cardiothorac Surg 1999;16:9-13. Crossref

5. Roques F, Michel P, Goldstone AR, Nashef SA. The logistic EuroSCORE. Eur Heart J 2003;24:881-2. Crossref

6. Nashef SA, Roques F, Sharples LD, et al. EuroSCORE II. Eur J Cardiothorac Surg 2012;41:734-44. Crossref

7. Chalmers J, Pullan M, Fabri B, et al. Validation of EuroSCORE II in a modern cohort of patients undergoing cardiac surgery. Eur J Cardiothorac Surg 2013;43:688-94. Crossref

8. Zheng Z, Li Y, Zhang S, et al. The Chinese coronary artery bypass grafting registry study: how well does the EuroSCORE predict operative risk for Chinese population? Eur J Cardiothorac Surg 2009;35:54-8. Crossref

9. Yap CH, Reid C, Yii M, et al. Validation of the EuroSCORE model in Australia. Eur J Cardiothorac Surg 2006;29:441-6. Crossref

10. Grant SW, Hickey GL, Dimarakis I, et al. Performance of the EuroSCORE models in emergency cardiac surgery. Circ Cardiovasc Qual Outcomes 2013;6:178-85. Crossref

11. Grant SW, Hickey GL, Dimarakis I, et al. How does EuroSCORE II perform in UK cardiac surgery; an analysis of 23 740 patients from the Society for Cardiothoracic Surgery in Great Britain and Ireland National Database. Heart 2012;98:1568-72. Crossref

12. Roques F, Nashef SA, Michel P, et al. Risk factors and outcome in European cardiac surgery: analysis of the EuroSCORE multinational database of 19030 patients. Eur J Cardiothorac Surg 1999;15:816-22. Crossref

13. Gogbashian A, Sedrakyan A, Treasure T. EuroSCORE: a systematic review of international performance. Eur J Cardiothorac Surg 2004;25:695-700. Crossref

14. Nashef SA, Roques F, Hammill BG, et al. Validation of European System for Cardiac Operative Risk Evaluation (EuroSCORE) in North American cardiac surgery. Eur J Cardiothorac Surg 2002;22:101-5. Crossref

15. Liu PH, Shih HH, Kang PL, Pan JY, Wu TH, Wu CJ. Performance of the EuroSCORE II model in predicting short-term mortality of general cardiac surgery: a single-center study in Taiwan. Acta Cardiol Sin 2022;38:495-503. Crossref

16. Shen L, Chen X, Gu J, Xue S. Validation of EuroSCORE II in Chinese patients undergoing coronary artery bypass surgery. Heart Surg Forum 2018;21:E036-9. Crossref

17. Zhang GX, Wang C, Wang L, et al. Validation of EuroSCORE II in Chinese patients undergoing heart valve surgery. Heart Lung Circ 2013;22:606–11. Crossref

18. Prince of Wales Hospital, The Chinese University of Hong Kong. Cardiac Surgery Report 2021–22. Available from: https://www.surgery.cuhk.edu.hk/cts/Cardiac_Surgery_Report_2021-22.pdf. Accessed 4 Sep 2024.

19. Kurniawaty J, Setianto BY, Widyastuti Y, Supomo S, Boom CE, Ancilla C. Validation for EuroSCORE II in the Indonesian cardiac surgical population: a retrospective, multicenter study. Expert Rev Cardiovasc Ther 2022;20:491-6. Crossref

20. Sembiring YE, Ginting A, Puruhito I, Budiono K. Validation of EuroSCORE II to predict mortality in post-cardiac surgery patients in East Java tertiary hospital. Med J Indones 2021;30:54-9. Crossref

21. Atashi A, Amini S, Tashnizi MA, et al. External validation of European System for Cardiac Operative Risk Evaluation II (EuroSCORE II) for risk prioritization in an Iranian population. Braz J Cardiovasc Surg 2018;33:40-6. Crossref

22. Zahara R, Soeharto DF, Widyantoro B, Sugisman, Herlambang B. Validation of EuroSCORE II scoring system on isolated CABG patient in Indonesia. Egypt Heart J 2023;75:86. Crossref

23. Musa AF, Cheong XP, Dillon J, Nordin RB. Validation of EuroSCORE II in patients undergoing coronary artery bypass grafting (CABG) surgery at the National Heart Institute, Kuala Lumpur: a retrospective review. F1000Res 2018;7:534. Crossref

24. Riley RD, Ensor J, Snell KI, et al. Calculating the sample size required for developing a clinical prediction model. BMJ 2020;368:m441. Crossref

25. Dendrite Clinical Systems. Databases for hospitals and clinics. Available from: https://www.e-dendrite.com/hospital-systems. Accessed 5 Sep 2024.

26. Paparella D, Guida P, Di Eusanio G, et al. Risk stratification for in-hospital mortality after cardiac surgery: external validation of EuroSCORE II in a prospective regional registry. Eur J Cardiothorac Surg 2014;46:840-8. Crossref

27. Stavridis G, Panaretos D, Kadda O, Panagiotakos DB. Validation of the EuroSCORE II in a Greek cardiac surgical population: a prospective study. Open Cardiovasc Med J 2017;11:94-101. Crossref

28. Nezic D, Spasic T, Micovic S, et al. Consecutive observational study to validate EuroSCORE II performances on a single-center, contemporary cardiac surgical cohort. J Cardiothorac Vasc Anesth 2016;30:345-51. Crossref

29. Garcia-Valentin A, Mestres CA, Bernabeu E, et al. Validation and quality measurements for EuroSCORE and EuroSCORE II in the Spanish cardiac surgical population: a prospective, multicentre study. Eur J Cardiothorac Surg 2016;49:399-405. Crossref

30. Koszta G, Sira G, Szatmári K, Farkas E, Szerafin T, Fülesdi B. Performance of EuroSCORE II in Hungary: a single-centre validation study. Heart Lung Circ 2014;23:1041-50. Crossref

31. Barili F, Pacini D, Capo A, et al. Does EuroSCORE II perform better than its original versions? A multicentre validation study. Eur Heart J 2013;34:22-9. Crossref

32. Biancari F, Vasques F, Mikkola R, Martin M, Lahtinen J, Heikkinen J. Validation of EuroSCORE II in patients undergoing coronary artery bypass surgery. Ann Thorac Surg 2012;93:1930-5. Crossref

33. Luo HD, Teoh LK, Gaudino MF, Fremes S, Kofidis T. The Asian system for cardiac operative risk evaluation for predicting mortality after isolated coronary artery bypass graft surgery (ASCORE-C). J Card Surg 2020;35:2574-82. Crossref

34. Ilcheva L, Risteski P, Tudorache I, et al. Beyond conventional operations: embracing the era of contemporary minimally invasive cardiac surgery. J Clin Med 2023;12:7210. Crossref

Periprosthetic joint infection (PJI) is an uncommon but severe complication of total knee arthroplasty (TKA). The existing literature indicates that approximately 1% to 2% of patients undergoing primary arthroplasty experience PJI; moreover, PJI is the leading cause of revision arthroplasty.1 2 Individuals with PJI may experience a substantial decrease in quality of life and must undergo complex and costly treatments to resolve this complication.3

A previous study at our institution, Queen Mary Hospital in Hong Kong, examined 2543 patients who underwent elective primary TKA between 1993 and 2013.4 During that period, the reported incidence of PJI was 1.34% and the most common causative organism was methicillin-sensitive Staphylococcus aureus (MSSA).4 The number of TKAs performed at our centre is rapidly increasing. In the past 8 years, clinicians at our institution have performed 85% of the total number of TKAs between 1993 and 2013, a span of 20 years. Considering the rapid population ageing in Hong Kong, we anticipate a continued increase in the number of TKAs. In recent years, various measures have been proposed to further reduce the incidence of PJI, including restrictions on blood transfusion rates, preoperative optimisation of modifiable risk factors, and the implementation of stringent culture techniques to improve microbial yield.5 6 7 However, the limited availability of local data makes it difficult to assess the effectiveness of these techniques in reducing PJI incidence. It is important to analyse the efficacy of these interventions as part of ongoing efforts to improve surgical outcomes at our centre. Furthermore, the increasing consumption of antibiotics over the past two decades has led to inevitable changes in the microbiological landscape of infectious organisms.8

This study had three objectives. First, it aimed to provide current local data on the incidence of PJI after elective primary TKA. Second, it sought to identify changes in the microbiological landscape of PJI; this information may guide future treatment and prevention strategies. Third, it aimed to showcase the efficacy of interventions to reduce PJI incidence and encourage their adoption beyond our institution.

Considering the measures introduced to reduce infection at our institution, we hypothesised that the incidence of PJI after primary TKA decreased over the past decade. We also hypothesised that the proportion of methicillin-resistant S aureus (MRSA)–related PJI increased during this period due to the increasing global consumption of antibiotics.

This retrospective cohort study compared the incidence and bacteriology of PJI after TKA at our institution. Participants were included if they underwent primary elective TKA at our institution between 2014 and 2021, and met the 2011 Musculoskeletal Infection Society (MSIS) criteria for PJI.9 Exclusion criteria were infection after revision arthroplasty, knee arthroplasty for malignant joint conditions, and active bacteraemia. The primary outcomes of interest were the incidence and bacteriological patterns of PJI. Secondary outcomes included preoperative patient demographics and time to onset of PJI.

The Hong Kong Hospital Authority’s Clinical Data Analysis and Reporting System and the Local Joint Replacement Registry were utilised to identify all TKAs performed at our institution between 2014 and 2021. Records were then searched using the keywords ‘orthopaedic aftercare’ and ‘periprosthetic joint infection’ to identify potential cases of PJI. Patients who did not meet the 2011 MSIS criteria for PJI were excluded; the remaining patients comprised the study cohort.10 Two senior authors of this study (PK Chan and KY Chiu) independently screened the patient database using the 2011 MSIS criteria to identify suitable patients for further data collection. Any uncertainties or disagreements were resolved through discussion.

Using a predefined data extraction form, the same two senior authors extracted the following data from the records of all included patients: intraoperative joint fluid culture results, age, sex, medical co-morbidities (eg, diabetes mellitus, rheumatoid arthritis, and immunosuppression), date of the index operation, surgical technique, operative time, date of re-operation, and postoperative antibiotic regimen.

A previous retrospective cohort study by Siu et al4 assessed the incidence and bacteriology of PJI among patients who underwent TKA at our institution between 1993 and 2013. From that study, we extracted data on the incidence and bacteriology of PJI, patient demographics, and time to onset of infection for comparison with our cohort. For both cohorts, we recorded the mean operative time, number of joint specialists involved, surgical technique, use of patient-specific instrumentation, and infection control protocols; these data were used to identify potential confounders that could influence the incidence of PJI.

Patients in our cohort were classified according to the time to infection onset as early, delayed, or late. Early-onset PJI was defined as infection occurring within 3 months of the index operation. These infections commonly arise from intraoperative contamination by highly virulent microorganisms and therefore constitute a key focus of intervention. Delayed-onset PJI was defined as infection occurring between 3 and 24 months after the index operation. These infections are also typically acquired during surgery but involve less virulent microorganisms. Late-onset PJI was defined as infection occurring over 24 months after surgery. These infections are often caused by haematogenous pathogens unrelated to the index operation.11

In accordance with our institution’s guidelines, patients were invited to attend follow-up at 2 weeks, 3 months, 6 months, and 12 months postoperatively. Patients without complications were subsequently scheduled for annual follow-up. Regarding infection control, the preoperative, perioperative, and postoperative protocols for elective primary TKA remained consistent throughout the study period in both cohorts. Intravenous antibiotic prophylaxis (1 g of cefazolin, or vancomycin for patients with a penicillin allergy) was administered 1 hour prior to skin incision. Intraoperatively, laminar airflow and body exhaust systems were utilised. Antibiotic-loaded cement was not routinely used, and a single postoperative wound management and rehabilitation programme was implemented throughout the study period. Postoperative antibiotics were not routinely administered.

Categorical variables were grouped for analysis; prevalence was calculated and group differences were tested with the Chi squared test. Continuous variables were compared using independent two-tailed t tests. A P value <0.05 was considered statistically significant. All statistical analyses were conducted using SPSS (Windows version 27.0; IBM Corp, Armonk [NY], United States).

In total, 2543 and 2171 primary TKAs were performed at our institution between 1993-20134 and 2014-2021, respectively. The incidence of PJI was 0.64% (n=14; 95% confidence interval=0.39-0.89) between 2014 and 2021, significantly lower than the 1.34% (n=34; 95% confidence interval=0.97-1.71) recorded between 1993 and 2013 (P=0.018).4

The mean age of the 14 patients with PJI in our cohort was 68.5 ± 7 years (range, 56-85). Of these patients, eight were men (57.1%) and six were women (42.9%). In terms of medical co-morbidities, seven patients had diabetes mellitus (50.0%), one had rheumatoid arthritis (7.1%), and one had end-stage renal disease requiring immunosuppression (7.1%). The mean follow-up period in our cohort was 4 years 9 months (interquartile range [IQR], 4 years 0 months to 6 years 11 months). There were no significant differences in age, sex distribution, or medical co-morbidities (diabetes mellitus and rheumatoid arthritis) between the two cohorts. The cohort demographics are compared in Table 1.4

We analysed other potential confounding factors (eg, mean operative time, number of joint specialists involved, and surgical approach) to minimise their effects on the primary and secondary outcomes.12 The indications for TKA did not change at our institution during the two time periods. Our institutional guidelines state that patients with Kellgren and Lawrence Grade 3 or 4 end-stage knee osteoarthritis and debilitating symptoms refractory to nonoperative treatment are candidates for TKA. The mean operative times for primary elective TKA were 1 hour 56 minutes during 1993-20134 and 1 hour 33 minutes during 2014-2021. The difference was not statistically significant (P=0.170). The number of joint specialists involved increased from four to six between the two periods. From 1993 to 2019, all TKAs were exclusively performed using the conventional approach.

After the computed tomography–based robotic arm–assisted system for total joint arthroplasty was introduced in 2019,13 surgeons at our institution could choose between robotic-assisted TKA and conventional TKA. Currently, there are no specific indications for either approach; the choice remains a matter of surgeon preference.14 To our knowledge, no studies have compared PJI incidence between robotic-assisted and conventional TKA; future research should explore the infection rate associated with each procedure. Patients undergoing robotic-assisted TKA at our institution follow the same postoperative protocol established for those undergoing conventional TKA (follow-up at 2 weeks, 3 months, 6 months, 12 months, and annually thereafter). Because robotic-assisted TKA was recently introduced at our institution, the mean follow-up duration for patients treated with this approach was short (14 months; IQR, 4.5-26).

Early-onset PJI occurred in 7.1% (n=1) of patients in the study cohort, arising 60 days after arthroplasty. In the 1993-2013 cohort, 29.4% (n=10) of patients experienced early-onset infection at a median of 17 days after arthroplasty (IQR, 9-32).4 However, the incidence of early-onset PJI did not differ significantly between the two cohorts (P=0.095) [Table 1]. Delayed-onset PJI occurred in 28.6% (n=4) of patients in the study cohort, occurring at a median of 6 months after arthroplasty (IQR, 5-7). Late-onset PJI occurred at a median of 3 years after arthroplasty (IQR, 2 years 1 month to 3 years 7 months). A larger proportion of patients experienced infection during the first year after surgery in the 1993-2013 cohort4 compared with the 2014-2021 cohort (59% vs 36%).

Methicillin-sensitive S aureus remained the most common causative organism in cases of PJI between 2014 and 2021, affecting 57.1% (n=8) of patients. The proportion of PJI cases caused by MSSA was significantly greater in the 2014-2021 cohort than in the 1993-2013 cohort, in which 26.5% (n=9) of patients were infected with MSSA (P=0.043) [Table 2].4

Methicillin-resistant S aureus was the second most common causative organism in cases of PJI between 1993 and 2013 (17.6%, n=6)4; Streptococcus spp. (14.3%, n=2) was the second most common causative organism between 2014 and 2021. The two cases of streptococcal infection in the 2014-2021 cohort comprised one with Streptococcus dysgalactiae and one with Streptococcus agalactiae. Other causative organisms in PJI cases within the 2014-2021 cohort included MRSA (7.1%, n=1), methicillin-sensitive coagulase-negative staphylococci (7.1%, n=1), and Escherichia coli (7.1%, n=1). Methicillin-resistant strains accounted for 40% of all staphylococcal infections between 1993 and 20134; this proportion was 11.1% between 2014 and 2021. Table 2 compares the microbiological patterns of PJI between the two cohorts.

There was a non-significant decrease in the proportion of patients with culture-negative PJI between the two cohorts, from 23.5% (n=8) between 1993 and 20134 to 7.1% (n=1) between 2014 and 2021 (P=0.186) [Table 2].

This study showed that the incidence of PJI after primary TKA at our institution significantly decreased. Worldwide, the reported incidence of PJI after primary elective TKA ranges from 1% to 2%.1 2 Over the years, our institution has implemented various measures to reduce the incidence of PJI after TKA, including a preoperative patient optimisation programme and a restrictive blood management programme. These measures are summarised in Table 3.5 6

The association between blood transfusion and increased perioperative morbidity in patients undergoing TKA is well documented.15 16 The American College of Surgeons National Surgical Quality Improvement Program reported that patients receiving transfusions experienced up to a tenfold increase in the risk of adverse postoperative outcomes.17 Based on these findings, a more restrictive transfusion approach has been implemented in the past several years after the 2015 study17 to improve postoperative outcomes. A retrospective study of 12 590 patients demonstrated significant decreases in complications, 30-day readmissions, and hospital length of stay following implementation of a patient blood management programme for patients undergoing prosthetic joint arthroplasty.18 The programme aimed to reduce transfusion requirements by optimising red cell mass, minimising blood loss, and defining appropriate indications for transfusion.18 A patient blood management programme was introduced at our institution in 2014; subsequently, the mean transfusion rate among patients undergoing TKA decreased from 31.3% in 2013 to 1.9% in 2018.6

The preoperative optimisation programme at our institution emphasises the optimisation of modifiable risk factors for PJI prior to TKA.5 Rheumatological diseases such as rheumatoid arthritis, juvenile inflammatory arthritis, ankylosing spondylitis, and psoriatic arthritis are known to increase the risk of PJI.19 20 A previous review of 2543 TKAs showed that the incidence of PJI was 3.1% in patients with rheumatoid arthritis, significantly higher than the 1.2% observed in patients without rheumatoid arthritis.4 At our institution, patients with rheumatoid arthritis who exhibit a persistently elevated erythrocyte sedimentation rate or C-reactive protein level are referred to a rheumatologist for further assessment and treatment prior to surgery. Diabetes mellitus is also strongly associated with an increased risk of PJI. All patients scheduled for elective TKA at our institution undergo universal glycated haemoglobin and fructosamine screening, with referral to an endocrinologist for optimisation if the glycated haemoglobin level exceeds 7.5%.21 Other modifiable risk factors monitored at our institution include weight control, vitamin D status, and nutritional status.22

From 2014 to 2021, MSSA was the most common causative organism in cases of PJI at our institution (57.1%, n=8). This finding is consistent with the existing literature, which indicates that S aureus is the most common causative organism in PJI after primary joint arthroplasty (19%-29% of cases worldwide).23 24 25 The unique virulence factors of S aureus enhance its ability to adhere to implants, facilitating aggressive biofilm formation and enabling replication and survival within this microenvironment.26 27

Owing to the increasing use of antibiotics in the community, we hypothesised that the number of antibiotic-resistant causative organisms would increase in our cohort of patients with PJI. The SENTRY Antimicrobial Surveillance Program evaluated 20-year trends in antimicrobial susceptibility among S aureus isolates across 427 centres in 45 countries.28 The authors reported that the prevalence of MRSA peaked at 44.2% in 2005-2008, then declined to 42.3% in 2009-2012 and 39.0% in 2013-2016.28 The incidence of MRSA among PJI cases in our study was consistent with findings from other regions. For example, a multicentre study in New Zealand showed that 9.1% of PJIs were attributable to MRSA.29

It is well established that early-onset infection typically occurs during surgery through intraoperative contamination, whereas late-onset PJI commonly arises from haematogenous spread.10 We observed a decrease in the proportion of early-onset infection between the two cohorts; however, this difference was not statistically significant (P=0.095). Additional measures should be implemented to further reduce the incidence of PJI after primary TKA at our institution.

In recent years, the incidence of culture-negative PJI has increased among patients undergoing total joint arthroplasty.30 This increase has been hypothesised to result from a higher prevalence of low-virulence organisms, premature antibiotic treatment, and failure to use enriched culture media.31 32 An inability to identify causative organisms in cases of PJI represents a serious problem for surgeons and infection control teams because of the uncertainties associated with antimicrobial selection. To reduce the incidence of culture-negative PJI, our institution implemented recommendations published by Tan et al,7 including extending the incubation period, using blood culture bottles and flasks, and collecting an adequate number of separate intraoperative tissue samples from patients with suspected PJI. The incidence of culture-negative PJI at our institution declined; however, this decline was not statistically significant (P=0.186).

This study had several limitations. First, it included patients treated at a single academic centre in Hong Kong; therefore, the findings may not accurately reflect changing trends in PJI incidence and bacteriology across the region. Further multicentre studies are warranted to better understand these trends. Second, the duration of patient recruitment differed between the present and former studies (7 years vs 20 years). Third, the number of patients varied between the two cohorts (2543 vs 2171). Despite this difference, we proceeded with the 7-year recruitment period considering the clinical value of evaluating recent PJI incidence and bacteriology in Hong Kong. Considerable effort was made to standardise patient characteristics and baseline co-morbidities to ensure comparability between cohorts. Fourth, given the relatively short mean follow-up duration (4 years 9 months), some cases of late-onset PJI may have occurred after completion of follow-up. Fifth, inconsistencies in record-keeping over the past decade prevented analysis of all documented risk factors for PJI; this limitation was unavoidable because of the retrospective study design. Sixth, the limited number of PJI cases hindered further subgroup analyses (ie, assessment and comparison of PJI incidence between conventional and robotic-assisted approaches). A similar study with a larger cohort may therefore be beneficial. Despite these limitations, the present study represents the largest series of PJI cases in Hong Kong to compare bacteriological patterns across two time periods. We believe that the findings have important clinical implications for PJI management in local hospitals.

This is the first study in Hong Kong to assess changes in the incidence and microbiological patterns of PJI after TKA across two time periods. Our findings have substantial clinical implications, as they demonstrate the effectiveness of interventional measures implemented at our institution in reducing the incidence of PJI, the rate of culture-negative PJI, and the number of early-onset cases. Prevention of PJI improves patient outcomes and reduces the economic burden on the healthcare system. Larger, multicentre, prospective studies are required to further elucidate bacteriological trends in PJI in Hong Kong.

Concept and design: All authors.
Acquisition of data: JR Khoo.
Analysis or interpretation of data: JR Khoo, PK Chan.
Drafting of the manuscript: JR Khoo.
Critical revision 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 manuscript was presented as an oral presentation at the 42nd Annual Congress of the Hong Kong Orthopaedic Association (5-6 November 2022, Hong Kong).

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

This research was approved by the Institutional Review Board of The University of Hong Kong/Hospital Authority Hong Kong West Cluster, Hong Kong (Ref No.: UW 25-585). The requirement for informed patient consent was waived by the Committee due to the retrospective nature of the research.

1. Pulido L, Ghanem E, Joshi A, Purtill JJ, Parvizi J. Periprosthetic joint infection: the incidence, timing, and predisposing factors. Clin Orthop Relat Res 2008;466:1710-5. Crossref

2. Zimmerli W, Trampuz A, Ochsner PE. Prosthetic-joint infections. N Engl J Med 2004;351:1645-54. Crossref

3. Premkumar A, Kolin DA, Farley KX, et al. Projected economic burden of periprosthetic joint infection of the hip and knee in the United States. J Arthroplasty 2021;36:148-49.e3. Crossref

4. Siu KT, Ng FY, Chan PK, Fu H, Yan CH, Chiu KY. Bacteriology and risk factors associated with periprosthetic joint infection after primary total knee arthroplasty: retrospective study of 2543 cases. Hong Kong Med J 2018;24:152-7. Crossref

5. Chan VW, Chan PK, Fu H, et al. Preoperative optimization to prevent periprosthetic joint infection in at-risk patients. J Orthop Surg (Hong Kong) 2020;28:2309499020947207. Crossref

6. Chan PK, Hwang YY, Cheung A, et al. Blood transfusions in total knee arthroplasty: a retrospective analysis of a multimodal patient blood management programme. Hong Kong Med J 2020;26:201-7. Crossref

7. Tan TL, Kheir MM, Shohat N, et al. Culture-negative periprosthetic joint infection: an update on what to expect. JB JS Open Access 2018;3:e0060. Crossref

8. Roberts SC, Zembower TR. Global increases in antibiotic consumption: a concerning trend for WHO targets. Lancet Infect Dis 2021;21:10-1. Crossref

9. Workgroup Convened by the Musculoskeletal Infection Society. New definition for periprosthetic joint infection. J Arthroplasty 2011;26:1136-8. Crossref

10. Parvizi J, Zmistowski B, Berbari EF, et al. New definition for periprosthetic joint infection: from the Workgroup of the Musculoskeletal Infection Society. Clin Orthop Relat Res 2011;469:2992-4. Crossref

11. Tande AJ, Patel R. Prosthetic joint infection. Clin Microbiol Rev 2014;27:302-45. Crossref

12. Wang Q, Goswami K, Shohat N, Aalirezaie A, Manrique J, Parvizi J. Longer operative time results in a higher rate of subsequent periprosthetic joint infection in patients undergoing primary joint arthroplasty. J Arthroplasty 2019;34:947-53. Crossref

13. The University of Hong Kong. HKUMed introduces the latest robotic arm assisted joint replacement technology for enhancing surgical precision [press release]. 2020 May 28. Available from: https://www.hku.hk/press/press-releases/detail/21111.html%2Dspecific%20planning. Accessed 1 Apr 2026.

14. Deckey DG, Rosenow CS, Verhey JT, et al. Robotic-assisted total knee arthroplasty improves accuracy and precision compared to conventional techniques. Bone Joint J 2021;103-B(6 Suppl A):74-80. Crossref

15. Everhart JS, Sojka JH, Mayerson JL, Glassman AH, Scharschmidt TJ. Perioperative allogeneic red blood-cell transfusion associated with surgical site infection after total hip and knee arthroplasty. J Bone Joint Surg Am 2018;100:288-94. Crossref

16. Lu Q, Peng H, Zhou GJ, Yin D. Perioperative blood management strategies for total knee arthroplasty. Orthop Surg 2018;10:8-16. Crossref

17. Ferraris VA, Hochstetler M, Martin JT, Mahan A, Saha SP. Blood transfusion and adverse surgical outcomes: the good and the bad. Surgery 2015;158:608-17. Crossref

18. Loftus TJ, Spratling L, Stone BA, Xiao L, Jacofsky DJ. A patient blood management program in prosthetic joint arthroplasty decreases blood use and improves outcomes. J Arthroplasty 2016;31:11-4. Crossref

19. Kunutsor SK, Whitehouse MR, Blom AW, Beswick AD; INFORM Team. Patient-related risk factors for periprosthetic joint infection after total joint arthroplasty: a systematic review and meta-analysis. PLoS One 2016;11:e0150866. Crossref

20. Morrison TA, Figgie M, Miller AO, Goodman SM. Periprosthetic joint infection in patients with inflammatory joint disease: a review of risk factors and current approaches to diagnosis and management. HSS J 2013;9:183-94. Crossref

21. Duensing I, Anderson MB, Meeks HD, Curtin K, Gililland JM. Patients with type-1 diabetes are at greater risk of periprosthetic joint infection: a population-based, retrospective, cohort study. J Bone Joint Surg Am 2019;101:1860-7. Crossref

22. Kong L, Cao J, Zhang Y, Ding W, Shen Y. Risk factors for periprosthetic joint infection following primary total hip or knee arthroplasty: a meta-analysis. Int Wound J 2017;14:529-36. Crossref

23. Triffault-Fillit C, Ferry T, Laurent F, et al. Microbiologic epidemiology depending on time to occurrence of prosthetic joint infection: a prospective cohort study. Clin Microbiol Infect 2019;25:353-8. Crossref

24. Zeller V, Kerroumi Y, Meyssonnier V, et al. Analysis of postoperative and hematogenous prosthetic joint-infection microbiological patterns in a large cohort. J Infect 2018;76:328-34. Crossref

25. Benito N, Franco M, Ribera A, et al. Time trends in the aetiology of prosthetic joint infections: a multicentre cohort study. Clin Microbiol Infect 2016;22:732.e1-8. Crossref

26. Kherabi Y, Zeller V, Kerroumi Y, et al. Streptococcal and Staphylococcus aureus prosthetic joint infections: are they really different? BMC Infect Dis 2022;22:555. Crossref

27. Sendi P, Rohrbach M, Graber P, Frei R, Ochsner PE, Zimmerli W. Staphylococcus aureus small colony variants in prosthetic joint infection. Clin Infect Dis 2006;43:961-7. Crossref

28. Diekema DJ, Pfaller MA, Shortridge D, Zervos M, Jones RN. Twenty-year trends in antimicrobial susceptibilities among Staphylococcus aureus from the SENTRY Antimicrobial Surveillance Program. Open Forum Infect Dis 2019;6(Suppl 1):S47-53. Crossref

29. Ravi S, Zhu M, Luey C, Young SW. Antibiotic resistance in early periprosthetic joint infection. ANZ J Surg 2016;86:1014-8. Crossref

30. Bejon P, Berendt A, Atkins BL, et al. Two-stage revision for prosthetic joint infection: predictors of outcome and the role of reimplantation microbiology. J Antimicrob Chemother 2010;65:569-75. Crossref

31. Berbari EF, Marculescu C, Sia I, et al. Culture-negative prosthetic joint infection. Clin Infect Dis 2007;45:1113-9. Crossref

32. Malekzadeh D, Osmon DR, Lahr BD, Hanssen AD, Berbari EF. Prior use of antimicrobial therapy is a risk factor for culture-negative prosthetic joint infection. Clin Orthop Relat Res 2010;468:2039-45. Crossref

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 old 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

Obstetric anal sphincter injury (OASIS) is a serious complication of vaginal delivery that can result in faecal incontinence, thereby impairing women’s quality of life. Reported prevalence rates of OASIS range from less than 1% to 11%.1 2 3 In the United Kingdom, the incidence tripled from 1.8% to 5.9% between 2000 and 2012, presumably due to improved detection techniques and increased awareness.4 In Hong Kong, the incidence increased from 0.04% in 2004 to 0.1% in 2009, and to 0.3% in 2014 during normal vaginal deliveries.5 Episiotomy, commonly performed during the second stage of labour to facilitate delivery and prevent excessive stretching of the perineal muscles, may increase intrapartum blood loss and perineal pain.6 The role of episiotomy in mitigating OASIS remains controversial.7 8 Consequently, the liberal use of episiotomy has declined in Hong Kong, with rates falling from 81% in 2004 to 66.2% in 2009 and 47.4% in 2014.5 Ethnic differences in pelvic floor biometry and pelvic organ mobility have been reported,8 9 and studies suggest that Asian women are more prone to OASIS.10 11 This study aimed to review the incidence of OASIS in our unit over the past decade in the context of declining episiotomy rates.

This study was conducted in Prince of Wales Hospital, a tertiary obstetrics and gynaecology unit with an annual delivery volume of approximately 4500 to 6000. All singleton vaginal deliveries—including spontaneous vaginal, ventouse, or forceps deliveries—between 1 January 2012 and 31 December 2021 were included. Breech and preterm deliveries were excluded. Maternal demographics were entered into the electronic record either antenatally by midwives or obstetricians if women had received antenatal care in our unit, or by midwives immediately after delivery. Maternal age and body mass index (BMI) were recorded at delivery. Macrosomia was defined as a birth weight of ≥4000 g. Most spontaneous vaginal deliveries were conducted by trained midwives or student midwives under supervision; instrumental deliveries were performed by trained obstetricians or trainees under senior supervision. When indicated, a left mediolateral episiotomy and a hands-on approach to protect the perineum were used by both midwives and doctors. Per vaginal and per rectal examinations were performed immediately after delivery. If OASIS was suspected, assessment was conducted by an obstetric specialist. The degree of OASIS was classified using the Abdul Sultan classification (Table 1).12 Delivery details were documented by midwives immediately after birth. Operative records for instrumental deliveries and OASIS repair, where applicable, were completed immediately after the procedure. Data were extracted from the hospital’s electronic delivery database between July 2022 and June 2023. Statistical analyses were performed using SPSS (Windows version 29.0; IBM Corp, Armonk [NY], United States). Descriptive analyses were used to examine demographics, mode of delivery, and the prevalences of episiotomy and OASIS. Means were compared between groups using the independent samples t test. Frequencies were compared using the Pearson Chi squared test or Fisher’s exact test, as appropriate. Trends were analysed using the Chi squared test for trend (Cochran–Armitage test). All risk factors were included in multivariable logistic regression analysis except epidural analgesia, nulliparity, and neonatal birth weight (justification provided in Results section). A P value of <0.05 was considered statistically significant.

A total of 43 732 deliveries were included in this study. The mean ± standard deviation maternal age at delivery was 31.5 ± 4.7 years and the median parity was 0 (interquartile range, 1). Of these, 22 566 (51.6%) were nulliparous and 21 166 (48.4%) were multiparous. Among the latter, 2268 (10.7%) had only previously delivered by Caesarean section and were therefore vaginally nulliparous. Data concerning previous delivery mode were missing for 905 women (4.3%). In total, 39 603 women (90.6%) had a normal vaginal delivery, 3528 (8.1%) had ventouse delivery, and 601 (1.4%) had a forceps delivery. Over the 10-year period from 2012 to 2021, the overall instrumental delivery rate and ventouse delivery rate declined significantly, from 13.2% to 12.0% (P<0.001) and from 11.8% to 8.6%, respectively (P<0.001) [Fig 1]. Overall, 23 325 women (53.3%) underwent episiotomy, whereas 20 407 (46.7%) did not; 326 women (0.7%) sustained OASIS, whereas 43 406 (99.3%) did not. The overall episiotomy rate decreased from 62.8% to 44.7% (P<0.001), with reductions observed in both nulliparous (from 89.2% to 68.5%; P<0.001) and multiparous women (from 31.7% to 23.8%; P<0.001). Conversely, the overall OASIS rate increased from 0.3% to 1.4% (P<0.001), with higher rates in nulliparous (from 0.4% to 2.5%; P<0.001) and multiparous women (0.1%-0.5%; P<0.001) [Fig 2].

The characteristics of the study population are summarised in Table 2. Episiotomy rates among women with and without OASIS were 51.8% and 53.3%, respectively (P=0.587). A higher proportion of women in the OASIS group were nulliparous (79.1% vs 51.4%; P<0.001) and vaginally nulliparous (85.9% vs 56.5%; P<0.001). Instrumental delivery was also more common in the OASIS group compared with the non-OASIS group (29.1% vs 9.3%; P<0.001). No statistically significant difference was observed between the type of instrumental vaginal delivery and the occurrence of OASIS (P=0.128). Women with OASIS had a lower BMI, a longer duration of labour, and delivered heavier neonates. No significant differences were observed in mean maternal age, ethnicity, gestational age, onset of labour, epidural analgesia, episiotomy, or macrosomia. All risk factors were included in the multivariable logistic regression analysis except epidural analgesia, nulliparity, and neonatal birth weight. Epidural analgesia was excluded because only one delivery with OASIS involved epidural analgesia, while nulliparity and neonatal birth weight were excluded due to their strong correlation with vaginal nulliparity and macrosomia, respectively. Macrosomia was considered to have greater clinical relevance than neonatal birth weight because a standard cut-off value exists. Multivariable logistic regression analysis revealed that vaginal nulliparity and instrumental delivery remained independent risk factors for OASIS, whereas BMI and labour duration did not. Induced labour (odds ratio [OR]=0.734, 95% CI=0.577-0.934; P=0.012) and episiotomy (OR=0.273, 95% CI=0.208-0.358; P<0.001) were identified as protective factors, while macrosomia (OR=2.754, 95% CI=1.435-5.284; P<0.001) was identified as a risk factor for OASIS (Table 3). Missing data were noted for BMI in 543 cases (1.2%) and for onset of labour in 82 cases (0.2%).

In the subgroup analysis of nulliparous women, the OASIS rate was significantly lower among those undergoing normal vaginal delivery with episiotomy compared to those without (0.6% vs 1.7%; P<0.001) and those undergoing instrumental delivery with episiotomy (1.7% vs 42.9%; P<0.001). Among multiparous women, the OASIS rate was significantly lower in those undergoing normal vaginal delivery without episiotomy (0.3% vs 0.5%; P=0.026) and those undergoing instrumental delivery with episiotomy (0.5% vs 23.5% without episiotomy; P<0.001). Among vaginally nulliparous women within the multiparous group, no statistically significant difference in OASIS rates was observed between normal vaginal deliveries with and without episiotomy; however, the OASIS rate was significantly lower among those undergoing instrumental deliveries with episiotomy compared with those without (0 vs 37.5%; P<0.001) [Table 4].

In recent years, many obstetric units in Hong Kong have promoted a reduction in episiotomy use in recent years. Our unit achieved substantial reductions in episiotomy rates among nulliparous and multiparous women between 2012 and 2021. Although the overall rate of OASIS remained low, considerable increases were observed in both groups during the study period. Vaginal nulliparity and operative vaginal delivery were identified as independent risk factors for OASIS, consistent with previous findings.7 11 Furthermore, episiotomy was identified as a protective factor against OASIS in multivariable logistic regression analysis (OR=0.273, 95% CI=0.208-0.358) [Table 3].

In nulliparous women, episiotomy was protective against OASIS in both normal and instrumental vaginal deliveries. These findings differ from those of previous large-scale studies.7 11 In a large retrospective study in the Netherlands involving over 281 000 vaginal deliveries,13 and in another study including more than 10 000 women in Australia,14 mediolateral episiotomy was shown to reduce the risk of OASIS in nulliparous women (OR=0.2113 and 0.54,14 respectively). However, Mahgoub et al11 in France reported no association between episiotomy and OASIS. In their cohort of 42 626 women, the overall OASIS rate was 1.2% and the overall episiotomy rate was only 10%.11 Perrin et al7 reported an episiotomy rate of 63.2% in nulliparous women and an OASIS rate of 0.7%, regardless of episiotomy use. In their analysis, episiotomy was not associated with OASIS in normal vaginal delivery but appeared to be protective in nulliparous women undergoing operative vaginal delivery at term.7

The above studies mainly involved women in Western populations. Several studies have indicated that Asian women have a two- to nine-fold increased risk of sustaining OASIS.15 16 17 18 19 In a study conducted in Israel involving over 80 000 women, including 997 of Asian origin, the OASIS rate among Asian women was 9 times higher than that among women of Western descent (3.5% vs 0.4%; P=0.001).16 Asian women also had a higher proportion of fourth-degree tears (17.1% vs 6.6%; P=0.039), despite smaller newborns (mean birth weight: 3318 g vs 3501 g; P=0.004).16 Anatomical differences between ethnic groups may contribute to this disparity. Cheung et al9 reported that pregnant women of East Asian origin had a thicker pubovisceral muscle, a smaller levator hiatus, and reduced pelvic organ mobility compared with pregnant women of Western descent. These factors may contribute to the higher risk of OASIS.9 Moreover, Bates et al20 found that a shorter perineal length measured during the second stage of labour prior to pushing was significantly associated with OASIS. Although a study conducted in Hawaii found no significant difference in perineal body length between Western and Chinese women, measurements were taken during the first stage of labour rather than before pushing.21 Further studies are needed to determine whether perineal body length differs during the second stage of labour. The reasons for the higher OASIS rates among Asian women remain unclear but are likely to be complex and multifactorial.

Another notable point is the higher rate of epidural analgesia use among Western women compared with Asian women (50%-90% vs 0%-2.2%), even within the same hospital setting where epidural analgesia is offered free of charge to all women.7 11 16 20 In the present study, the rate of epidural analgesia was low throughout the study period. In this cohort, epidural analgesia was not associated with OASIS. A meta-analysis examining risk factors for OASIS found no association with epidural analgesia; however, it included only two studies.22 In contrast, Mahgoub et al11 identified epidural analgesia as a protective factor for OASIS, whereas another meta-analysis reported it as a risk factor.19 These conflicting findings suggest that the role of epidural analgesia in OASIS remains unclear.

There is limited literature on the role of episiotomy in normal vaginal delivery among multiparous women. In the present study, episiotomy did not protect multiparous women from OASIS, except in the context of instrumental vaginal delivery. Indeed, episiotomy may increase the risk of OASIS in this group.23 However, we noted that episiotomy was protective against OASIS among multiparous women undergoing instrumental vaginal delivery (OR=0.028). This finding is supported by a Dutch study which reported five-fold and ten-fold reductions in OASIS during vacuum and forceps deliveries, respectively.24 In light of these findings, we recommend a more restrictive approach to episiotomy among multiparous women undergoing normal vaginal delivery.

The rising trend of OASIS over the past decade may also be attributable to improvements in clinical detection following the promotion of more thorough post-delivery assessments by both midwives and obstetricians. Kwok et al25 reported that the prevalence of occult OASIS—detected by endoanal ultrasound but not identified by clinical examination after delivery—was as high as 7.8% after normal vaginal delivery and 3.8% after instrumental delivery. Subsequently, regular OASIS workshops were introduced to train midwives and doctors in performing standardised vaginal and rectal examinations after vaginal delivery. When a major perineal tear is suspected, immediate reassessment by an obstetric specialist is conducted. This practice has been shown to improve the detection rate of OASIS.26 We also analysed trends in instrumental vaginal delivery over the 10-year period. Overall, decreasing trends were observed for both instrumental and ventouse deliveries. The rate of forceps delivery remained similar or showed a slight decrease, except in 2021. Therefore, the rising trend in OASIS is unlikely to be explained by changes in instrumental delivery rates.

The strengths of this study include its large sample size, 10-year study period, and the documented reduction in episiotomy rates, which allowed evaluation of the role of episiotomy in OASIS. Our unit is a tertiary centre with the highest delivery volume in Hong Kong, and this represents the largest retrospective study to date focusing on an Asian population. However, as a retrospective study, missing data were noted during data collection and entry. In addition, several risk factors previously identified in meta-analyses—such as the duration of the second stage of labour, fetal head position at delivery, history of previous OASIS, and shoulder dystocia—were not analysed in the present study,19 27 representing a key limitation. Furthermore, some cases of OASIS may have been missed on clinical examination. High-quality research is needed to further investigate OASIS, given its substantial impact on women’s quality of life.

With a substantial reduction in episiotomy rates, a corresponding increase in the rate of OASIS was observed. Episiotomy was protective against OASIS among nulliparous women undergoing singleton normal vaginal delivery and instrumental delivery. It also conferred protection in multiparous women undergoing instrumental delivery but not in those having normal vaginal delivery. Among vaginally nulliparous women within the multiparous group, the OASIS rate was significantly higher in those undergoing instrumental deliveries without episiotomy, similar to the rate observed in nulliparous women. Conversely, the OASIS rate was higher in the episiotomy group during normal vaginal delivery, although this difference was not statistically significant and may have been influenced by the small sample size. Further high-quality research is warranted, and women should be informed of these findings to enable informed decision-making regarding episiotomy.

Concept or design: SSC Chan, RYK Cheung.
Acquisition of data: SSC Chan, RYK Cheung, TW Chau, YY Lau, SM Ng, TM Tso.
Analysis or interpretation of data: SSC Chan, YY Lau.
Drafting of the manuscript: YY Lau.
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 Ms LL Lee, their research assistant, for her assistance with data acquisition, analysis, and interpretation.

Findings from this study were partially presented as an e-poster at the Royal College of Obstetricians and Gynaecologists World Congress 2024, Muscat, Oman, 15-17 October 2024.

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.: 2022.259). The requirement for patient consent was waived by the Committee due to the retrospective nature of the research. The study complied with the Declaration of Helsinki and the International Council for Harmonization Guideline for Good Clinical Practice.

1. Tung CW, Cheon WC, Tong WM, Leung HY. Incidence and risk factors of obstetric anal sphincter injuries after various modes of vaginal deliveries in Chinese women. Chin Med J (Engl) 2015;128:2420-5. Crossref

2. Jangö H, Langhoff-Roos J, Rosthøj S, Sakse A. Modifiable risk factors of obstetric anal sphincter injury in primiparous women: a population-based cohort study. Am J Obst Gynecol 2014;210:59.e1-6. Crossref

3. Hsieh WC, Liang CC, Wu D, Chang SD, Chueh HY, Chao AS. Prevalence and contributing factors of severe perineal damage following episiotomy-assisted vaginal delivery. Taiwan J Obstet Gynecol 2014;53:481-5. Crossref

4. Gurol-Urganci I, Cromwell DA, Edozien LC, et al. Third- and fourth-degree perineal tears among primiparous women in England between 2000 and 2012: time trends and risk factors. BJOG 2013;120:1516-25. Crossref

5. Hong Kong College of Obstetricians and Gynaecologists. Territory-wide Audit in Obstetrics & Gynaecology. 2014. Available from: https://www.hkcog.org.hk/hkcog/Download/Territory-wide_Audit_in_Obstetrics_Gynaecology_2014.pdf. Accessed 1 May 2020.

6. Woolley RJ. Benefits and risks of episiotomy: a review of the English-language literature since 1980. Part II. Obstet Gynecol Surv 1995;50:821-35. Crossref

7. Perrin A, Korb D, Morgan R, Sibony O. Effectiveness of episiotomy to prevent OASIS in nulliparous women at term. Int J Gynaecol Obstet 2023;162:632-8. Crossref

8. Abdool Z, Dietz HP, Lindeque BG. Ethnic differences in the levator hiatus and pelvic organ descent: a prospective observational study. Ultrasound Obstet Gynecol 2017;50:242-6. Crossref

9. Cheung RY, Shek KL, Chan SS, Chung TK, Dietz HP. Pelvic floor muscle biometry and pelvic organ mobility in East Asian and Caucasian nulliparae. Ultrasound Obstet Gynecol 2015;45:599-604. Crossref

10. Brown J, Kapurubandara S, Gibbs E, King J. The great divide: country of birth as a risk factor for obstetric anal sphincter injuries. Aust N Z J Obstet Gynaecol 2018;58:79-85. Crossref

11. Mahgoub S, Piant H, Gaudineau A, Lefebvre F, Langer B, Koch A. Risk factors for obstetric anal sphincter injuries (OASIS) and the role of episiotomy: a retrospective series of 496 cases. J Gynecol Obstet Hum Reprod 2019;48:657-62. Crossref

12. de Leeuw JW, Struijk PC, Vierhout ME, Wallenburg HC. Risk factors for third degree perineal ruptures during delivery. BJOG 2001;108:383-7. Crossref

13. Okeahialam NA, Taithongchai A, Thakar R, Sultan AH. The incidence of anal incontinence following obstetric anal sphincter injury graded using the Sultan classification: a network meta-analysis. Am J Obstet Gynecol 2023;228:675-88.e13. Crossref

14. Hauck YL, Lewis L, Nathan EA, White C, Doherty DA. Risk factors for severe perineal trauma during vaginal childbirth: a Western Australian retrospective cohort study. Women Birth 2015;28:16-20. Crossref

15. Grobman WA, Bailit JL, Rice MM, et al. Racial and ethnic disparities in maternal morbidity and obstetric care. Obst Gynecol 2015;125:1460-7. Crossref

16. Baruch Y, Gold R, Eisenberg H, et al. High incidence of obstetric anal sphincter injuries among immigrant women of Asian ethnicity. J Clin Med 2023;12:1044. Crossref

17. D’Souza JC, Monga A, Tincello DG. Risk factors for perineal trauma in the primiparous population during non-operative vaginal delivery. Int Urogynecol J 2020;31:621-5. Crossref

18. Yeaton-Massey A, Wong L, Sparks TN, et al. Racial/ethnic variations in perineal length and association with perineal lacerations: a prospective cohort study. J Matern Fetal Neonatal Med 2015;28:320-3. Crossref

19. Hu Y, Lu H, Huang Q, et al. Risk factors for severe perineal lacerations during childbirth: a systematic review and meta-analysis of cohort studies. J Clin Nurs 2023;32:3248-65. Crossref

20. Bates LJ, Melon J, Turner R, Chan SS, Karantanis E. Prospective comparison of obstetric anal sphincter injury incidence between an Asian and Western hospital. Int Urogynecol J 2019;30:429-37. Crossref

21. Tsai PJ, Oyama IA, Hiraoka M, Minaglia S, Thomas J, Kaneshiro B. Perineal body length among different racial groups in the first stage of labor. Female Pelvic Med Reconstr Surg 2012;18:165-7. Crossref

22. Barba M, Bernasconi DP, Manodoro S, Frigerio M. Risk factors for obstetric anal sphincter injury recurrence: a systematic review and meta-analysis. Int J Gynaecol Obstet 2022;158:27-34. Crossref

23. Eggebø TM, Rygh AB, von Brandis P, Skjeldestad FE. Prevention of obstetric anal sphincter injuries with perineal support and lateral episiotomy: a historical cohort study. Acta Obstet Gynecol Scand 2024;103:488-97. Crossref

24. van Bavel J, Hukkelhoven CW, de Vries C, et al. The effectiveness of mediolateral episiotomy in preventing obstetric anal sphincter injuries during operative vaginal delivery: a ten-year analysis of a national registry. Int Urogynecol J 2018;29:407-13. Crossref

25. Kwok SP, Wan OY, Cheung RY, Lee LL, Chung JP, Chan SS. Prevalence of obstetric anal sphincter injury following vaginal delivery in primiparous women: a retrospective analysis. Hong Kong Med J 2019;25:271-8. Crossref

26. Andrews V, Sultan AH, Thakar R, Jones PW. Occult anal sphincter injuries—myth or reality? BJOG 2006;113:195-200. Crossref

27. Pergialiotis V, Bellos I, Fanaki M, Vrachnis N, Doumouchtsis SK. Risk factors for severe perineal trauma during childbirth: an updated meta-analysis. Eur J Obstet Gynecol Reprod Biol 2020;247:94-100. Crossref