AI-Enabled Echocardiography Interpretation With Multitask Deep Learning—Reply

In Reply We appreciate the thoughtful comments from Ms Wu and colleagues and Dr Zhang and colleagues about our article. Both Letters comment on the representativeness of patient cohorts and whether demographic bias could emerge in model predictions. We agree that bias is a key consideration for the responsible deployment of AI algorithms, which we addressed with multisite validation across diverse cohorts and statistical analyses on group fairness. Of note, our cohorts spanned a large US-based health system representative of the local demographic census (19.1% non-White, 7.9% Hispanic), including an emergency department cohort with point-of-care imaging (34.2% non-White, 11.4% Hispanic), as well as geographically distinct cohorts from Europe and California (11.1% Hispanic). More importantly, the model’s robust performance across cohorts with variable demographic composition suggests that PanEcho, the AI system, learns key phenotypes through direct associations rather than confounders. Supplemental analysis also demonstrated that the model performed equitably across sex for 13 of 15 diagnostic tasks and across race for 25 of 26 groupwise comparisons. We therefore observed no evidence of demographic bias in the applications of the AI system, although continued evaluation is essential.