Vinny Jodoin
Research done in collaboration with Dr. Suzanne Lenhart of the Department of Mathematics, University of Tennessee Knoxville, USA and Dr. Heidi Hanson, of Oak Ridge National Laboratory, USA
Talk is at 11:00 AM Central Time (calculating your local time…)
Abstract
Heart disease and its associated comorbidities represent a leading global cause of morbidity and mortality.Yet the physiological signatures preceding clinical diagnosis remain poorly characterized. The mechanistic links between short-term changes in physiological state and long-term multimorbidity progression are similarly understudied. Early detection of disease-preceding signals could meaningfully improve diagnostic lead time enabling a fundamental shift from reactive diagnosis to proactive, individualized intervention. Passively collected heart rate (HR) data from consumer wearable devices offer a promising avenue for identifying these patterns, though inherent noise and individual variability require novel analytical approaches. This study leverages Fitbit HR data from over 500 participants in the National Institutes of Health’s All of Us Research Program to identify latent physiological states and characterize their relationship to long-term disease progression. Dynamic Time Warping (DTW) was applied to build a seven-day HR trajectories, followed by K-Medoids clustering to yield six reproducible clusters, or physiological state phenotypes. Transition dynamics were modeled as a Markov chain, revealing a stationary distribution with approximately time-reversible transitions.
To connect short-term deviations to long-term clinical outcomes, Group-Based Trajectory Modeling (GBTM) identifies distinct latent subgroups of multimorbidity progression across the observation period. Incorporating lagged cluster history improved next-state classification by approximately 47% over random chance, and cycle asymmetry in the transition structure indicates directional preference in physiological state progression: together suggesting non-random pathways between the systems states. These findings demonstrate the utility of extracting clinically meaningful phenotypes from large-scale wearable data and contribute to pathways for transitioning from reactive diagnosis toward proactive, individualized identification of disease trajectory transitions in diverse population cohorts.