Health status change detection using anomaly detection in latent spaces

Abstract

Individual-specific changes in health conditions are detected using a latent space mapping generated from baseline physiological data collected in a longitudinal study of the individual. Historical baseline data in an n-dimensional input space is modeled into a k-dimensional latent space, where k<n. The mapping for the model is then used to convert new physiological data into a latent space point, which can be used to determine if the associated newly collected physiological data is anomalous and thus indicative of a change in health condition. Anomalies can be inferred from poor fit between the point and baseline data within the latent space, or form large error between data reconstructed into the input space from the point and the original new physiological data.