Bridging the Explainability Gap

A Novel Framework for Interpretable Deep Learning in Clinical Diagnostics

Abstract

Deep learning models now match or exceed specialists on several diagnostic tasks, yet clinicians remain reluctant to rely on them because the reasoning behind a prediction is hidden. Simple models are transparent but less accurate, and the commonly used remedy, post-hoc saliency maps, often highlights regions that do not correspond to anything a clinician would recognise. This paper proposes a framework that narrows this explainability gap by building clinical meaning into the model rather than adding it afterwards. Using the electrocardiogram as a worked example, the framework learns a compact set of disentangled latent factors with a variational autoencoder whose capacity is controlled during training, links those factors to named waveform components such as the P wave, QRS complex, ST segment, and T wave, and feeds them to a diagnostic classifier whose decisions can be traced back to the factors.....