Improving Machine Learning Based Detection of Freezing of Gait Using Data Synthesis Methods

Abstract

Freezing of gait (FoG) is a common type of motor impairment in Parkinson's disease (PD) associated with falls. Pharmacologic treatments have failed to prevent FoG, resulting in a need for the development of non-pharmaceutical interventions. This study aims to use the ADAptive SYNthetic (ADASYN) sampling algorithm to improve the automatic detection of FoG episodes in people with PD. Eighteen PD patients performed a series of daily walking tasks in a lab environment, with nine experiencing FoG. We trained and evaluated different classifiers using signals collected from wearable accelerometers placed on the patients' ankles. The low ratio of FoG labeled data compared with normal gait created an imbalanced dataset. The analysis of results revealed that the sensitivity of the SVM model to FoG could be improved from 91.9% to 96.5% by creating new synthesized FoG instances near the boundaries of two classes (normal gait and FoG). The SVM model can detect FoG events with an average sensitivity and specificity of 91.5% and 95.2% trained by the original dataset, and 95.8% and 95.0% trained by synthesized balanced dataset in patient-dependent models. Also, ADASYN slightly improved the average F-score of patient-dependent models from 90.6% to 91.2%, while reducing patient-independent model performance from 83.4% to 79.8%.