Data Driven Disease Dynamics Models

Models that explicitly consider the dynamic nature of disease progression promise a more comprehensive analysis of longitudinal datasets and disease characterization. This paper presents a novel framework that utilizes optimal reaction coordinates (RCs) to describe disease progression as a diffusion on a free energy landscape. This method addresses key challenges, including the curse of dimensionality, irregular sampling, and data imbalance, providing a theoretically optimal representation of stochastic disease dynamics. Additionally, we introduce a new validation criterion that outperforms traditional metrics like AUC in distinguishing between optimal and sub-optimal RCs. Our approach offers a comprehensive and practical tool for analyzing disease dynamics, facilitating early diagnosis and targeted medical interventions.