Optimizing Insulin Dosing for Type 1 Diabetes with Thyroid Dysfunction Using Q-Learning: A Personalized Approach to Chronic Disease Management

Thyroid dysfunction frequently coexists with Type 1 Diabetes (T1D), creating complex clinical challenges due to the critical interplay between thyroid hormone fluctuations and insulin sensitivity. Existing insulin dosing protocols typically do not account for these dynamic comorbid interactions, often leading to suboptimal glycemic control and increased adverse event risk. To address this gap and prioritize the clinical interpretability necessary for adoption, we propose a novel Reinforcement Learning (RL) framework based on tabular Q-learning that explicitly models discrete thyroid dysfunction severity within the patient state and incorporates the delayed pharmacodynamic effects of thyroid medications into a dual-objective reward function. This deliberate design enables personalized, transparent insulin dosing policies that optimize both glycemic control and thyroid hormone stabilization. We evaluate our approach on the real-world T1DGranada cohort comprising adults with T1D and hypothyroidism. Our comorbidity-aware, interpretable model achieves a 15% improvement in Time-in-Range (TIR) and a 42% reduction in hypoglycemic events compared to standard clinical baselines, while also significantly enhancing thyroid hormone stabilization rates. Offline evaluation techniques including importance sampling and Fitted Q-Evaluation (FQE) validate the robustness and reliability of the learned policies. Furthermore, expert endocrinologist blind review confirms high clinical alignment with 83% agreement. These results underscore the importance of explicitly modeling multimorbidity and delayed treatment effects in interpretable RL frameworks to advance personalized chronic disease management and facilitate clinical trust and integration.