Weakly Supervised Representation Learning for Trauma Injury Pattern Discovery

Abstract

Given the complexity of trauma presentations, particularly in those involving multiple areas of the body, overlooked injuries are common during the initial assessment by a clinician. We are motivated to develop an automated trauma pattern discovery framework for comprehensive identification of injury patterns which may eventually support diagnostic decision-making. We analyze 1,162,399 patients from the Trauma Quality Improvement Program with a disentangled variational autoencoder, weakly supervised by a latent-space classifier of auxiliary features. We also develop a novel scoring metric that serves as a proxy for clinical intuition in extracting clusters with clinically meaningful injury patterns. We validate the extracted clusters with clinical experts, and explore the patient characteristics of selected groupings. Our metric is able to perform model selection and effectively filter clusters for clinically-validated relevance.