A deep learning model for assistive decision-making during robot-aided rehabilitation therapies based on therapists' demonstrations.
David Martínez-Pascual, José M Catalán, Luís D Lledó and 2 others
PMID 39891159WHAT IT FOUND
A robot model learned when one therapist helped eight patients during point-to-point arm exercises.
It timed assistance better than a rule that waited a set time, but it was not tested live or shown to improve recovery.
Key findings
01A single therapist's assistance was recorded as force on the robot handle, and trials were labelled assisted when force exceeded 0.5 N.
02After fine-tuning, the model classified 38 of 52 unassisted trajectories and 32 of 40 assisted trajectories correctly for two test participants.
03The fine-tuned model had a lower false positive rate than the time-dependent method, 0.2692 versus 0.7885.
STILL TO COME
How it was doneWhat they foundWhat it means for PTs
Read the rest of this summary
You get three full summaries a month, free, and we do not ask for a card. Search, the TL;DRs and your library stay unlimited either way.
What it does not show
Only one therapist provided the assistance demonstrations, so the model learned that one person's timing, not a range of therapist judgement. Six participants' data were used for training and validation, and only two participants' data were used for testing. The model was evaluated on recorded trajectories, and real-time use was not tested. The main outcome was agreement with the therapist's assistance decision, not patient movement recovery or function. The test set was small and had few assisted trajectories, with 40 assisted and 52 unassisted trajectories. The method was designed and evaluated only for point-to-point exercises.
Declared interests
The supplied publication metadata lists non-U.S. Government research support, but the article text does not include a competing interests statement.
The easy way to misread this
Do not read the model's accuracy as evidence that robot-assisted therapy improves recovery. It was trained and tested offline on recorded trajectories from a small sample, with one therapist's assistance decisions, and real-time use was not evaluated.