Context-aware optimization of follow-up intervals for type 2 diabetes care using Markov decision processes

Chronic disease management relies on regular patient-provider interactions to monitor progression and control. For type 2 diabetes (T2D), current guidelines prescribe fixed intervals between primary care visits, overlooking heterogeneity in clinical trajectories and patient characteristics. This study formulates subpopulation-specific follow-up planning as a context-specific Markov decision process (MDP) to optimize follow-up intervals using electronic health records from 17,499 patients across 10 primary care clinics. Contexts were identified through (i) principal component analysis for dimensionality reduction of health-trajectory variables and (ii) clustering patients using the resulting components and patient-level features. The MDP used a factorized transition model with state components including hospitalization, observed HbA1c, measurement recency, and HbA1c trend and was solved using approximate dynamic programming. Policies were evaluated against benchmarks using fitted Q evaluation (FQE). Two contexts emerged, representing lower- and higher-risk subpopulations. The framework recommended shorter intervals as HbA1c increased, shifting to 3-month follow-up for moderately elevated values and 1-month follow-up for very high values. The high-risk subpopulation shifted to shorter intervals at lower HbA1c levels than the low-risk subpopulation. Across the full cohort and both subpopulations, the derived policies outperformed benchmarks. These findings indicate that the framework has the potential to improve subpopulation-specific follow-up planning in primary care.