Towards AI-based precision rehabilitation via contextual model-based reinforcement learning.

No real stroke survivors were treated; all outcome data were computer-simulated. The patient model was simplified and based on a chronic-stroke model, not acute or subacute recovery. The simulated outcome was the subjective Motor Activity Log, not an objective functional measure such as the Action Research Arm test. The dose recommendations were simplified to hours only and did not target specific upper-limb tasks. The simulation assumed no measurement noise and a fully observed patient state, which is not true in clinic. The model update method is computationally intensive and may not scale to large datasets without approximation methods. The authors state that a large and variable real rehabilitation database is needed to train and validate the framework before clinical use.

The supplied text names the National Institutes of Health as funder.

Do not read the simulation results as evidence that this AI system improves patient outcomes in clinic. It tested computer-generated patients, not real people, and the authors say validation with large real data is needed before adoption.