Digital Twin Workflow for Continuous Glucose Forecasting: Real-Time Data Assimilation Using Reinforcement Learning

Digital twins for physiological forecasting require robust mechanisms for continuously assimilating streaming data and adapting to patient-specific dynamics. Threshold-triggered model updates can restore predictive performance after substantial drift, but their reactive nature limits continuous adaptation. We propose a proactive digital twin framework that couples a pretrained PatchTST foundation model with an Actor-Critic reinforcement learning (RL) agent for residual correction. Using a 30-minute look-back window and recent residual history, the agent generates bounded corrections ({+/-} 0.05) to the first-step baseline forecast, with corrected predictions recursively propagated over a 15-step horizon. We evaluate both an RL-only workflow and a hybrid workflow combining RL correction with threshold-triggered fine-tuning using minute-by-minute continuous glucose monitoring data from 10 critically ill patients with sepsis. Across all patients, the pretrained RL agent improves forecasting accuracy without patient-specific retraining. Both workflows reduce trajectory-averaged mean absolute percentage error, ensemble relative error, and maximum errors compared with the uncorrected baseline and previously developed reactive digital twin models. These findings support residual reinforcement learning as a lightweight mechanism for continuous, real-time data assimilation that preserves the foundation model's global predictive prior while enabling patient-specific forecast adaptation, with potential applications in critical care decision support.