Deep reinforcement learning finds optimal retraining strategies for patients with knee osteoarthritis

Reinforcement learning (RL) can be used to optimise motion with respect to important biomechanical variables by searching over the space of kinematics and muscle excitations in predictive simulation. We used kinematic data obtained from patients with medial compartment osteoarthritis (OA) of the knee to initialise patient-specific models of motor control. From this starting point, kinematic constraints were removed and further training was performed to reduce the medial knee joint contact force, a well-established predictor of clinical severity. The final learned kinematic adaptations demonstrated features that accord with the targets of established physical retraining strategies but varied across patients, highlighting the utility of this personalised approach.