Automatic therapy planning based on concept drift.

Only three healthy adults took part, so the study cannot speak to whether the system helps anyone recover; the authors say validating it in stroke survivors is still the essential next step. The real sessions followed a fixed game schedule set by the data-collection protocol, which forced some transitions between games that did not reflect the system's reasoning or the participant's actual state. How much the system values improvement versus overload was set by ranking clinical priorities with the collaborating physiotherapist rather than from measured patient outcomes, and the game order, the available actions and the assumptions about how patients respond to harder tasks come from clinical judgement and from healthy adults; a clinician with different priorities would get different recommendations. The physiotherapist who helped set those priorities also took part in judging the recommendations, and was the only expert involved, so the 87.5% agreement reflects one clinician's judgement and may be inflated. The assumption about how likely a patient is to improve after a change in difficulty has not been checked against the responses of stroke patients. Continuous movement data had to be broken into categories to fit the model, which loses some detail.

The authors state that financial support was received for this work and/or its publication. Specifically, CONAHCYT granted one author an academic scholarship for MSc studies in Computer Science. No company that sells the technology is named as a funder. Separately, the physiotherapist who helped set the system's reward priorities also took part in evaluating its recommendations, which the authors identify as a possible source of bias.

Do not read the 87.5% agreement as evidence this system improves arm recovery. It was tested on three healthy adults rather than stroke survivors, and the physiotherapist who agreed with its choices had also helped set the priorities the system was built to follow.