Multimodal engagement estimation in paediatric robot-assisted gait training: integrating physiological sensing and biomechanical interaction.

Only 20 children took part, all with cerebral palsy, and there was no comparison group of typically developing children, so it is not known whether the same signals would behave the same way in other populations or settings. The data were split into 30-second windows rather than by child, so windows from the same child could end up in both the training and the testing data. The reported accuracy is therefore likely to be better than the model would manage on a child it has never seen. Engagement was rated by one clinician watching the sessions. No second rater was used, so there is no measure of how consistently those ratings could be reproduced, and the model is essentially learning to copy one person's judgement. No session was recorded without the Lokomat's augmented feedback games, so the study cannot say what part those games play in engagement. The classifiers were compared descriptively. No statistical test was run, so the differences in accuracy between models are not confirmed differences. The children ranged from GMFCS level I to level V, and the authors note the sample was too small to check whether the model performs as well at every severity level. In a single pooled model, children with severe impairment may produce different baseline signals from those with mild impairment. Facial temperature around the mouth and nose can also be changed by breathing, physical effort, fatigue and orofacial problems, so it is not a clean measure of engagement on its own. The analysis was done offline; the time the system takes to produce an answer in real time was not measured. Global explanations of the model were used. No per-patient or per-window explanations were produced.

Funded by the Italian Ministry of University and Research through the European Union's NextGenerationEU programme. The text supplied contains no declaration of competing interests.

Do not read these figures as a monitoring device you could use in clinic. The model was reproducing one clinician's engagement ratings in an offline dataset, and windows from the same child appeared in both the training and the testing data, so the accuracy is likely better than it would be with a child the model has not seen. Do not conclude either that robot training or feedback games increase engagement; the study was not designed to test that.