Despite great promise, AI systems have yet to initiate major changes in the daily practice of medicine. This is largely due to several crucial unmet needs of doctors and healthcare providers. With the goal of developing clinically-useful medical-AI, we systematically address these needs to arrive at system for electrocardiogram analysis which provides accurate and interpretable predictions while also being highly transparent regarding its limitations.

Our system presents major steps towards addressing the following unmet needs: (1) prediction explainability conforming to medical practice; (2) uncertainty estimation while upholding statistical requirements; (3) rejection of samples for which the model is irrelevant; (4) natural handling of data containing unknown conditions;  and (5) generalization across patients.

We demonstrate applicability by simulating the screening of a large population under real-world settings, while adhering to rigorous statistical requirements necessary for clinical deployment.