Prediction and detection of freezing of gait in Parkinson's disease from plantar pressure data using long short-term memory neural-networks.

Only 11 male participants were tested, and only 7 froze; Participant 7 contributed 221 of 362 freeze episodes. The study was a single laboratory session with added cognitive and motor tasks to provoke freezes, not a real-world assessment. The reported outcomes were model classification metrics, not patient outcomes such as falls, mobility, or independence. The model misclassified many standing periods as freezes: 65.3% of standing data were false positives. Prediction used a binary setup that grouped pre-freeze and freeze together, so it did not test a separate pre-freeze warning class. Freeze labels were identified by study authors, with a second rater consulted only if uncertain. The insoles used were wired, single-use laboratory sensors, not a finished wearable product.

Funded by Microsoft Canada, Waterloo Artificial Intelligence Institute, Network for Aging Research, University of Waterloo, and Natural Sciences and Engineering Research Council of Canada. No author conflicts-of-interest declaration is provided in the supplied article text.

Do not conclude that plantar-pressure insoles can detect freezing of gait reliably in clinics. The model detected 343 of 361 lab freeze episodes in only seven participants, and 65.3% of standing data were false positives, so it is not ready for patient care.