Parkinson's Disease Detection from Gait Force Signals Using Machine Learning

RESEARCH AND TEACHING ONLY: NEVER FOR DIAGNOSIS, SCREENING OR TREATMENT. NOT A MEDICAL DEVICE. Can under-foot force from walking separate people with Parkinson's disease from controls, and what does a fair test show? On PhysioNet gaitpdb (162 IDs = 154 persons of three studies), split by person, logistic regression on four stride measures reaches AUC 0.827, Random Forest on 26 features 0.865, a 1-D CNN on raw force 0.801 and stopwatch gait speed alone 0.763. The paired intervals show no clear gain of the larger models over the four measures, and the four measures are not clearly better than gait speed. A random split of windows, labelled what not to do, makes the same CNN look 99.7 % accurate.

Category: Biomedical Engineering

Biomedical Engineering Project BM07 from ProTechSC Academy, with graphical abstract, video demonstration and project guidance.