Freezing of Gait Classification in Parkinson’s Patients Using Machine Learning

Abstract

This study aims to design personalized machine learning models for the classification of time-series data to detect freezing of gait (FoG) in Parkinsons’ patients in short time intervals. FoG is the medical terminology for sudden episodes of an inability to move in patients that suffer from Parkinson’s disease. Data collected experimentally by Xuanwu Hospital were used. Each 0.002-second interval is labeled as FoG positive or negative by physicians. Using information gain statistics, it was determined that out of 58 features of accelerometer, EEG, and EMG measurements, 35 measurements provide the most information about FoG presence. Features were normalized via z-score normalization. For feature vectors, data are grouped into 0.5-second batches with .002 second timeframes for LSTM; while data are grouped into 0.5-second intervals for other models. The FoG positive/negative classes were balanced through SMOTE. All models were hyperparameter trained through 10-fold cross-validation. The F-1 scores of LSTM, Random Forest, SVM, Decision Tree, and Logistic Regression are 89.71%, 89.69%,  87.00%, 74.44%, 67.21% respectively. Of the models analyzed, LSTM has the highest recall at 93.16%, while Random Forest has the highest precision at 94.34%. LSTM detects the most positive instances, while Random Forest has precise detection. LSTM has a higher F-1 score, indicating it is better at balancing precision and recall. These personalized short interval-input models can be implemented in wearable devices to detect freezing of gait to aid physicians’ assessment of disease severity and treatment.