PCPs: Patient Cardiac Prototypes to Probe AI-based Medical Diagnoses, Distill Datasets, and Retrieve Patients

Clinical deep learning systems often generate population-based and opaque medical diag-noses. This is in contrast to how primary care physicians make decisions, often adapting population-based protocols to the unique patient under consideration. Inspired by the work-flow of such physicians, we develop a framework for learning embeddings, referred to as patient cardiac prototypes (PCPs), which capture information that is unique to an individual patient’s electrocardiogram (ECG) data. Through rigorous evaluation on three publicly-available ECG datasets, we show that PCPs allow researchers to inspect why a particular diagnosis was made. We also demonstrate that PCPs are effective dataset distillers, where they can be used to train a model in lieu of a dataset orders of magnitude larger to achieve comparable performance. We show that PCPs can also be exploited to retrieve similar patient data across clinical databases. Our framework contributes to the development of transparent and patient-specific clinical deep learning systems.