Freezing of gait assessment with inertial measurement units and deep learning: effect of tasks, medication states, and stops.
Po-Kai Yang, Benjamin Filtjens, Pieter Ginis and 5 others
PMID 38350964WHAT IT FOUND
A deep learning model using leg sensors matched expert ratings of freezing of gait severity in Parkinson's disease.
It worked well for standard walking and turning tests, but only if trained on the specific task being tested. It also distinguished freezing from voluntary stops if stops were included in training.
Key findings
01The model showed strong agreement with clinical experts in calculating percentage of time frozen (ICC = 0.92) and number of freezing episodes (ICC = 0.95) during standardized tests.
02Models trained on one specific task (like turning in place) failed to accurately detect freezing in a different task (like walking and turning), indicating that training data must match the test task.
03Including voluntary stops in the training data allowed the model to distinguish freezing from stopping, whereas a model trained without stops misidentified 35.23% of stops as freezing.
STILL TO COME
How it was doneWhat they foundWhat it means for PTsWhat it means for OTs
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
Small sample size of only 12 patients, all with severe freezing (at least one episode per day). The study used simulated free-living conditions (stopping during lab tasks) rather than actual daily life data. The model overestimated freezing severity compared to experts, partly by labeling shuffling and festination as freezing. Inter-rater agreement between experts was not calculated, so the model's accuracy was judged against a single consensus annotation. It is unknown if the model works for patients with mild freezing or other manifestations like akinetic freezing.
Declared interests
The authors declared no conflicts of interest. Funding was provided by KU Leuven internal funds and scholarships.
The easy way to misread this
Do not assume this technology is ready for home monitoring. The study showed that models trained only on lab tasks with stops still misidentified 7.61% of stops as freezing, and the model's performance dropped significantly when tested on tasks it hadn't seen. Current tools require task-specific calibration and cannot yet reliably distinguish freezing from normal pauses in complex daily activities.