A multimodal machine learning approach to forecast upper limb motor recovery after stroke using kinematic and electromyographic data - A pilot-study.

The study included only 11 stroke survivors, so the results cannot be expected to generalise to typical caseloads. Most participants received standard therapy plus robot-assisted training, and one participant received standard therapy only, so the contribution of any single component cannot be separated. The models were evaluated only by cross-validation within the same small dataset, with no independent patient sample. The anomaly-detection model was trained on only seven recoverers, and the authors note that many test values were close to the classification threshold. The grouping of recoverers and non-recoverers was based on before and after Fugl-Meyer scores, which may bias the anomaly-detection results. Difficulty reaching targets may reflect unfamiliarity with the robot, not motor impairment, because familiarization was not assessed.

Do not read the high classification score as proof that this system can choose a patient's therapy. It was tested in 11 subacute stroke survivors using only internal cross-validation, and most participants received standard therapy plus robot training, so the contribution of any component cannot be separated.