Explore the scholarly record.

Abstract

Abstract Background Gender differences in emergency care are well documented and generally disadvantage women, including at the triage stage. AI-based triage tools may reproduce or amplify these differences. We previously developed a model identifying, among individuals seeking emergency or unscheduled care through a medical call centre, those requiring referral to an emergency department (ED), and observed better discrimination in men than in women. This study aimed to analyse the effect of the sex composition of the training data on this gender difference in discriminatory performance. Methods We used telephone calls (from 2018 to 2023) and medical records from SOS Médecins Grand Paris (France), which provides same-day home visits for patients. At the end of each visit, the doctor records the patient's outcome (left at home versus referred to EDs; 3% referrals) in the patient's medical file. Calls were transcribed using an automatic speech recognition model and then processed using a text‑based model to classify whether the patient required ED attendance. From sex-balanced training, validation and test subsets (n=601 560), three subsets of identical size (n=300 780) were derived - female, male, and sex-balanced-reduced-by-half - and used to train three models. Discrimination was compared by sex using receiver operating characteristic area under the curve (ROC-AUC). Results Of 1 876 172 calls followed by a home visit, 1 098 748 (60%) were matched with their medical record. Each model was trained on 164 980 calls (10% referrals) and evaluated on 67 900 validation and 67 900 test calls (3% referrals). The female model achieved a ROC-AUC of 0.812 (95% confidence intervals (CI) 0.803–0.820) among women, and the male model 0.835 (95%CI 0.827–0.843) among men. For all three models, discrimination was higher among men than among women, significantly so for the male and the sex-balanced-reduced-by-half models, but not for the female model. Conclusions Balancing training sets by sex is neither a necessary nor a sufficient condition for fairness. A male-only model generalised poorly to women, whereas a female-only model appeared applicable to both sexes, suggesting structural differences that warrant placing fairness considerations at the design stage rather than treating them as secondary. Trial registration ClinicalTrials.gov Identifier: NCT07428109 (registered on 23 February 2026)