Adaptive Shared-Basis Learning for Subject-Specific Density Estimation

Abstract

Dense repeated measurements from wearable devices and sensing technologies provide rich information about individual behavioral and physiological patterns, but common summaries or smoothed trajectories may miss distributional features relevant to health outcomes. For example, individuals with similar mean glucose may differ substantially in variability or exposure to extreme levels. Subject-specific densities capture such differences, but commonly used methods estimate each subject’s density separately and therefore do not learn or exploit structure shared across subjects. Additionally, measurement error in wearable recordings can further obscure latent distributions. We develop a density function neural network for estimating subject-specific densities while learning shared structure across subjects. Applications to continuous glucose monitoring and wearable physical activity data illustrate how subject-specific densities can provide information beyond conventional summaries for health prediction and association analyses.