Stability-aware LSTM-MPC framework for personalized levothyroxine dosing in Hashimoto's thyroiditis

Hypothyroidism, most commonly caused by Hashimoto's thyroiditis, is characterized by insufficient production of thyroid hormones and is typically treated with oral levothyroxine (LT4). However, determining patient-specific LT4 dosage remains challenging due to significant inter-individual variability and the need for repeated clinical adjustments. In this study, a stability-aware, data-driven dosing framework is developed for personalized LT4 therapy. The proposed approach combines a Long Short-Term Memory (LSTM) neural network for predicting free thyroxine (FT4) concentrations with a model predictive control (MPC) strategy for dose optimization. To ensure safe and reliable operation, Input-to-State Stability (ISS) conditions are incorporated within the control framework, ensuring boundedness of the system states and stability of the closed-loop system. The method is evaluated using simulated patients generated with Thyrosim, a validated mathematical simulator of thyroid hormone regulation, under varying biological parameters. Simulation results demonstrate that the proposed framework regulates FT4 concentrations toward patient-specific euthyroid levels during the initial phase of treatment and maintains stable hormone levels throughout the simulation period. Furthermore, consistent performance across a heterogeneous population of simulated patients highlights the robustness of the approach. These findings suggest that the proposed ISS-LSTM-MPC framework provides a promising foundation for automated, personalized LT4 dosing and represents a step toward clinically applicable decision-support systems for thyroid hormone therapy.