Markovian Representation Learning for Longitudinal Electronic Health Records

Abstract Deep learning models for electronic health records are typically trained with token-level reconstruction or task-specific supervision, encouraging contextual predictability rather than explicit modeling of temporal dynamics. We propose a Markovian representation learning framework that treats longitudinal EHR data as realizations of a latent stochastic process and trains embeddings to admit low-complexity transition operators in latent space. By optimizing transition likelihood alongside information-preserving objectives, the learned representations approximate sufficient statistics of underlying clinical state, yielding compact dynamical modes with interpretable time scales. Across multiple prediction tasks and external hospital transfer settings, Markov-structured pretraining improves discrimination, temporal forecasting performance, and robustness relative to supervised and masked-event baselines. We further demonstrate that the learned transition structure supports LLM-assisted summarization and agentic cohort exploration, enabling clinician-friendly interpretation of latent progression patterns. These results suggest that explicitly modeling temporal transition structure provides a principled complement to token-centric foundation modeling for longitudinal medical data.