fetching data ... 
Background: Sex and gender differences are crucial for investigating differences in health-related outcomes in general population and provision of equitable care [1, 2]. Rheumatoid arthritis (RA) is a relatively common autoimmune rheumatic disease, characterised by sex bias in prevalence and outcomes.
Objectives: To assess the acknowledgment and mitigation of sex bias within studies using supervised machine learning (ML) for improving clinical outcomes in RA.
Methods: Design: A systematic review of original studies published in English between January 2018 and November 2023 using the PUBMED and EMBASE databases has been conducted (PROSPERO ID CRD42023431754).
Study Selection Criteria : original research papers, use of supervised ML to predict clinically relevant outcomes in RA, and publication within the specified date interval.
Data Extraction and Synthesis : Papers were scored on whether they reported, attempted to mitigate, or successfully mitigated various types of bias: training data bias, test data bias, input variable bias, output variable bias, and analysis bias. The quality of ML research in all papers was assessed to evaluate the robustness of the data and the overall rigor of the research using six quality metrics with equal weighting (scored 0,1 or 2), thus giving a total quality score out of 12. Papers scoring 0-3 were deemed low quality, 4-8 medium and 9-12 considered high quality.
Results: Out of N=52 studies eligible for inclusion for qualitative synthesis and analysis, all but one (which included only 17% females) had a female skew in the study participants (across all studies, the mean proportion of females was 74±12%), reflecting the female sex bias in RA prevalence in general population. The following outcomes have been explored using ML strategies: prediction of treatment response (N=23 papers) and disease activity scores (N=11 papers), improvement in diagnostic accuracy (N=10 papers), assessing the joint damage (N=10 papers); and identification of distinct patient subgroups (N=3 papers). The ML algorithms employed by eligible papers were: Random Forest (N=30 papers), Regression (N=25 papers), Neural Networks (N=21 papers), Support Vector Machine (N=17 papers), Boosted Tree (N=15 papers), K Nearest Neighbours (N=6 papers), Naïve Bayes (N=4 papers), Other (3 papers, including Gaussian Process, Pathway Supported Models and Hidden Markov Models). The majority of papers (42/52 papers, 80.7%) did not acknowledge any potential sex bias (scored zero). The remaining ten papers scored only 1 each across all bias categories. Three papers assessed bias in model performance by sex-disaggregating their results, one paper acknowledged potential sex bias in input variables, and six papers in their output variables, predominantly disease activity scores. No paper attempted to mitigate any type of sex bias. Only one paper was scored as high quality, N=38 papers were determined as medium and N=13 as low quality. The stand-out anomaly was the lack of sample size description, as it was the case with N=42/52 papers. There was no correlation between the quality of the papers and their consideration of sex bias (r = 0.06, p = 0.65), with the only high-quality paper scoring zero for sex bias consideration.
Conclusion: Our analysis demonstrates that sex bias is generally not considered in the development of ML models in RA. These findings raise important questions about how relevant the results of these papers are for clinical application and highlight the lack of contribution of these studies to mitigation of sex bias in healthcare. Increased consideration of the types of bias outlined and investigated in this systematic review is imperative to ensuring the inclusion of diverse and representative data in research, aspects which are ultimately critical for the development of equitable and accurate models to enhance the fairness and reliability of ML applications in healthcare.
REFERENCES: [1] Mauvais-Jarvis F, Bairey Merz N, …, Sandberg K, Suzuki A. Sex and gender: modifiers of health, disease, and medicine. Lancet. 2020 Aug 22;396(10250):565-582.
[2] Peng J, Jury EC, Dönnes P, Ciurtin C. Machine Learning Techniques for Personalised Medicine Approaches in Immune-Mediated Chronic Inflammatory Diseases: Applications and Challenges. Front Pharmacol. 2021 Sep 30;12:720694. doi: 10.3389/fphar.2021.720694.
Acknowledgements: “This research was conducted and supported by Haleon. We extend our gratitude to Haleon’s partnership with Women in Data® and Juliette Scott for facilitating the collaboration between industry and academia.”
Disclosure of Interests: None declared.
© The Authors 2025. This abstract is an open access article published in Annals of Rheumatic Diseases under the CC BY-NC-ND license (