PhyMod: A generative model of longitudinal physiology for interpretable disease risk

Chronic diseases arise from complex interactions among lifestyle, environmental, and physiological factors that evolve over time. Here, we present PhyMod, a longitudinal model of physiology that learns interpretable, multivariable signatures from routine clinical measurements and lifestyle exposures to predict the incidence of chronic disease. By modeling variable trajectories, PhyMod also forecasts future physiological states and disease risk for up to 10 years in the future. These learned signatures provide an interpretable representation of the physiological patterns underlying disease progression. Using two randomized clinical trials of lifestyle intervention in type 2 diabetes, we further show that PhyMod captures intervention-associated shifts in physiology consistent with observed trial outcomes.