Real-Time Disease Detection with Energy-Based Models

Abstract

With increasing numbers of common diseases affecting the general population, diagnoses are becoming more time-consuming, expensive, and stressful for patients, requiring frequent laboratory tests and specialized equipment to track disease progression. Recent innovations in wearable medical sensing and machine learning have enabled portable smartwatch-based \textit{health decision support systems} (HDSSs). HDSSs use real-time samples of physiological signals to proactively diagnose diseases through deep neural network classifiers, then alert medical professionals so that further laboratory tests can be administered once a detection result is positive. However, these models have a large parameter space, placing a latency and energy burden on battery-powered smartwatches and causing high model variance (i.e., sensitivity to the specific training and validation datasets used before deployment). Neural network classifier models are also generally poor at uncertainty estimation, making them difficult for detecting disease comorbidities or distinguishing between mutually exclusive disease outcomes. This work uses dataset preprocessing, along with state-of-the-art energy-based model frameworks including Joint Energy-Based Models (JEM) and Variational Entropy Regularized Approximate (VERA) maximum likelihood generation, to optimize the accuracy, variance, size, latency, and confidence calibration of existing detection model frameworks across multiple disease categories, helping to enhance their usefulness in pervasive healthcare applications.