An Interpretable Workflow for Cohort Phenotyping Using Longitudinal Wearable Data.

Abstract

Consumer wearable devices non-invasively capture continuous physiological and behavioral signals, positioning them as promising tools for precision medicine applications, in particular disease diagnostics. Nevertheless, while most of the literature focuses on models and algorithms for disease prediction or classification, there is limited reporting on data distributions, heterogeneity, and subgroups within cohorts. The common practice of collapsing rich longitudinal data into single summary metrics obscures latent phenotypes and potential biomarkers and risks imposing mismatched models on the data. To address this gap, we introduce an interpretable workflow for cohort-level phenotyping that leverages longitudinal timestamped wearable metrics, clinically relevant risk strata, and unsupervised machine learning methods. Using two real-world cohorts as examples, the workflow identified three sleep phenotypes and three physio-behavioral phenotypes (characterized as low-, intermediate-, and high-risk). Across cohorts, these phenotypes showed modest but consistent separation, reflecting partially overlapping but systematically different profiles. The first two principal components explained 74-77% of the variance for the sleep phenotypes and roughly 60% for the physio-behavioral phenotypes, with the component loadings recurring across independent cohorts. By formalizing an exploratory analysis for wearable cohort data, this workflow improves transparency and cross-study comparability, supports clinically interpretable subgroup discovery, and provides a principled foundation for downstream digital health modeling.