Data-Driven Prediction of Freezing of Gait Events From Stepping Data

Freezing of gait (FoG) is typically a symptom of advanced Parkinson's disease (PD) that negatively influences the quality of life and is often resistant to pharmacological interventions. Novel treatment options that make use of auditory or sensory cues might be optimized by prediction of freezing events. These predictions might help to trigger external sensory cues—shown to improve walking performance—when behavior is changed in a manner indicative of an impending freeze (i.e., when the user needs it the most), rather than delivering cue information continuously. A data-driven approach is proposed for predicting freezing events using Random Forrest (RF), Neural Network (NN), and Naive Bayes (NB) classifiers. Vertical forces, sampled at 100 Hz from a force platform were collected from 9 PD subjects as they stepped in place until they at least had one freezing episode or for 90 s. The F1 scores of RF/NN/NB algorithms were computed for different IL (input to the machine learning algorithm), and GL (how early the freezing event is predicted). A significant negative correlation between the F1 scores and GL, highlighting the difficulty of early detection is found. The IL that maximized the F1 score is approximately equal to 1.13 s. This indicates that the physiological (and therefore neurological) changes leading to freezing take effect at-least one step before the freezing incident. Our algorithm has the potential to support the development of devices to detect and then potentially prevent freezing events in people with Parkinson's which might occur if left uncorrected.