Prediction of Freezing of Gait in Parkinson's Disease Using Time-Series Data from Wearable Sensors

Freezing of gait is a common symptom in patients with Parkinson's disease, which is mainly manifested by sudden interruption of walking, so that patients cannot walk forward effectively. Freezing of gait is associated with increased risk for falls, loss of independence, and impaired quality of life. Prediction of freezing of gait before it occurs can not only greatly improve the effect of intervention, but also avoid the occurrence of freezing to some extent. In this study, we propose an autoregressive predictive model to predict feature time series, followed by support vector machine classifier to discriminate freezing from gait sequence based on the predicted features. After extracting highly relevant features from accelerometer and gyroscope signals, we perform feature selection using correlation and maximal information coefficient to reduce the feature list to 13 candidates. The approach results in an average precision of81.61 %, recall of83.37%, accuracy of85.08%, with an expected prediction horizon of 1.63 s. Our gait freezing prediction algorithm yields consistent results across different subjects and has certain academic and clinical application significance.