Vision-based assessment of parkinsonism and levodopa-induced dyskinesia with pose estimation.
Michael H Li, Tiago A Mestre, Susan H Fox and 1 others
PMID 30400914WHAT IT FOUND
In nine people with Parkinson’s disease and troublesome dyskinesia, a camera-based system estimated dyskinesia severity from videos of selected tasks.
It worked better for talking tasks than drinking tasks and poorly identified normal movement, so it is not ready for clinical use.
Key findings
01The pose-estimation features predicted dyskinesia severity from communication videos with a mean correlation of 0.661 with clinician ratings, but predicted severity from drinking videos with a mean correlation of 0.043.
02In multiclass classification of communication videos, the system was correct 71.4% of the time overall, but it detected normal movement with only 9.4% sensitivity.
03Predicting total validated scores for the dyskinesia and parkinsonism scales gave correlations of 0.741 and 0.530 with clinician ratings.
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 nine participants were studied, all with idiopathic Parkinson's disease and stable bothersome peak-dose dyskinesia; there were no healthy control participants. The participants were selected for moderate dyskinesia, so the results do not show how the system performs across the full range of Parkinson's motor complications. The videos were not originally recorded for computer-vision analysis, so camera angle, zoom, camera shake, and severe occlusions required manual cleaning and segmentation. Pose estimation used low-resolution resized frames, and knee tracking was poor because hospital gowns provided little texture. Face dyskinesia subscores were not modelled, and head turning in the communication task was tracked separately from the pose skeleton. The toe tapping method relied on assumptions about foot position and was not validated for foot motion. Binary thresholds were chosen to balance classes, not to match clinical definitions of normal versus abnormal movement. Performance was evaluated with repeated cross-validation in the same small sample, not with an independent validation dataset. Some clinical scale items were omitted, including dressing and rigidity, so predictions were based on a subset of the full assessment. The study did not establish what level of agreement with clinician ratings would be clinically useful, nor did it test whether the system detects important changes better than clinicians.
Declared interests
Funding was provided by the Natural Sciences and Engineering Research Council of Canada, the Toronto Rehabilitation Institute, and the Toronto General and Western Hospital Foundation. The supplied text does not include an author conflict-of-interest statement.
The easy way to misread this
Do not read the 71.4% overall accuracy or the better agreement with clinician ratings as evidence this can replace clinician assessment. It was tested in nine people with moderate dyskinesia, had no healthy control group, and detected normal movement in only 9.4% of normal cases.