Predicting clinically significant motor function improvement after contemporary task-oriented interventions using machine learning approaches.

This was a secondary analysis of data from earlier trials, not a new prospective prediction study. The models were developed and tested within the same dataset, with only a small test sample and no external validation in new patients. The test sample had far fewer low responders than high responders, which may affect specificity, especially for the artificial neural network model. Predictions were made for change immediately after intervention, not for lasting improvement at follow-up. Only upper-limb motor function was predicted, not other outcomes such as quality of life, participation, or activities of daily living. Only a limited set of machine learning algorithms was tested, so another algorithm may perform differently. Patients were selected for moderate to mild hemiparesis and had to be able to follow study procedures, so the models may not apply to more severe or cognitively impaired stroke patients. The interventions were delivered by trained occupational therapists in research settings, which may not match routine clinical practice. The authors recommend recording three baseline items, but they did not test whether using them changes patient selection or outcomes.

Funding was provided by Chang Gung Memorial Hospital, Linkou; the Healthy Aging Research Center at Chang Gung University under the Featured Areas Research Center Program within the Higher Education Sprout Project by the Ministry of Education in Taiwan; and the National Health Research Institutes in Taiwan. No other author conflict statement is given in the supplied text.

Do not read this as a ready clinical decision tool or as evidence that one task-oriented therapy works better than another. The models were built from pooled trial data and tested on a small internal sample, so they need external validation before use in patient care.