Personalized prophylactic therapy optimization in hemophilia A using a hybrid PK-PD-TTE model and deep RL.

Hemophilia A is a genetic bleeding disorder caused by a deficiency or absence of Factor VIII, leading to recurrent and spontaneous hemorrhages. Standard treatment typically involves regular prophylactic infusions of clotting factors to prevent bleeding episodes. However, individual variations in treatment response and reliance solely on plasma Factor VIII levels provide an imprecise assessment of bleeding risk. This study presents an intelligent dose-control system utilizing Deep Reinforcement Learning (DRL) integrated with a hybrid Pharmacokinetic-Pharmacodynamic Time-to-Event (PK-PD-TTE) environment. Within this framework, the control policy is learned using a Deep Q‑Network (DQN), enabling the agent to adapt treatment decisions through interaction with the simulated physiological environment. Unlike conventional methods, this framework allows the agent to observe continuous physiological states via Endogenous Thrombin Potential (ETP) and learn optimal dosing policies through trial-and-error. The proposed DQN agent was benchmarked against standard prophylaxis, a Fuzzy Logic controller, and a Bayesian Adaptive Model-Informed Precision Dosing (MIPD) strategy. Simulation results from a virtual cohort (N = 200) demonstrate that the DQN agent achieves a safety profile comparable to Bayesian MIPD while significantly improving factor utilization efficiency. Notably, in patients with low bleeding‑risk phenotypes, the DRL‑based approach achieved a similar bleeding rate to MIPD while reducing annual factor VIII consumption by approximately 39% relative to MIPD and by up to 70% compared to the reference prophylaxis protocol adopted as the simulation baseline. These findings suggest that incorporating reinforcement learning with mechanistic PD feedback can act as a powerful complementary layer to established clinical protocols, facilitating personalized and cost-effective hemophilia management.