Machine learning-based study of the predictors of clinically important change in patient-reported outcomes in bilateral upper-limb function in patients receiving robotic stroke rehabilitation.

The sample was small for machine learning, with 123 participants and only 25 in the testing set. The study was a secondary analysis, so it may have missed important predictors such as self-efficacy or rehabilitation psychology. A single train-test split was used, and the authors noted this may limit model learning and increase risk of overfitting. All participants received the same bilateral robotic-primed program, so the findings do not show how other therapies or components compare. The results may not generalize to other cultures or to people with very severe arm impairment, because participants had baseline FMA-UE scores above 10. The two ABILHAND cutoffs produced different top predictors, so the choice of cutoff changes the answer.

The work was funded in part by the Ministry of Science and Technology under grant numbers MOST 111-2314-B-002-168-MY3 and NSTC-113-2811-B-002-075. The authors declared that financial support was received for the work or its publication.

Do not read the 0.76 and 0.80 accuracy values as proof that these models can choose patients for robotic therapy. The testing set had only 25 participants, and the authors called for further validation in larger samples.