Evaluating sex bias in artificial intelligence-based echocardiographic analysis: a retrospective study on individuals...

Artificial intelligence (AI) has the potential to improve echocardiography. However, AI systems may exhibit sex bias, reproducing societal stereotypes and performing worse for certain groups, potentially leading to unequal healthcare outcomes. Therefore, the aim of this study was to investigate whether there are sex-related differences in the agreement between a fully automated AI-based analysis and routine echocardiographic report generated by the interpreting cardiologist in patients undergoing transthoracic echocardiography.

This retrospective study evaluated seven echocardiographic parameters in 95 women and 88 men by comparing AI-generated values with clinical assessments. Participants included individuals with normal cardiac function (34 women, 31 men) and patients with systolic (30 women, 29 men) or diastolic dysfunction (31 women, 28 men). Parameters included ejection fraction, left ventricular volume, left atrial volume, é velocities (septal and lateral), tricuspid valve regurgitation, and E/é ratio. Potential sex-based differences were assessed by comparing agreement between clinical and AI-generated values in women and men using correlations, absolute errors, Bland-Altman analyses and AUC values.

Correlation coefficients in women ranged from 0.67 to 0.93 and in men from 0.32 to 0.93. Absolute errors did not differ significantly between sexes for five of the seven evaluated parameters. However, men showed a higher absolute error for TR velocity (0.15 vs. 0.07, p = 0.001), whereas women showed a higher absolute error for the E/é ratio (0.76 vs. 0.39, p < 0.001). Bland-Altman analysis showed larger mean differences in men for five of seven parameters. The magnitude of the differences in AUC values between women and men was small, ranging from 0.017 (TR velocity) to 0.048 (left ventricular ejection fraction).

This study did not demonstrate systematic sex-related differences in AI performance, but the findings require confirmation in larger cohorts. Continued monitoring for bias remains important as AI integration in echocardiography expands.