Automatic therapy planning based on concept drift

Static rehabilitation protocols often struggle to keep pace with the dynamic, non-linear reality of human motor recovery. Traditional therapy models frequently assume stationarity, relying on fixed performance schemes that fail to capture the highly individualized nature of human recovery and skill acquisition. This paper proposes an adaptive decision-support framework that integrates concept drift detection within Bayesian Networks to drive a Markov Decision Process (MDP) for automated therapy planning. By continuously monitoring real-time data streams, the system can identify both parametric and structural shifts in a patient's motor control during virtual rehabilitation sessions. These detected drifts serve as signal for the MDP, which searches and recommends an therapeutic adjustment. As a proof of concept, the framework was evaluated using data from healthy participants interacting with the platform and assessed through expert clinical feedback. In this preliminary evaluation, the generated policies balanced engagement with progression, rewarding improvement while penalizing task overload and frustration. While validation in clinical populations remains as future work, the integration of probabilistic tracking and sequential decision-making offers a promising basis for personalized, data-driven virtual rehabilitation in which therapeutic adjustments can evolve alongside the user.