Abstract
Freezing of gait (FOG) is a motor symptom experienced by people with Parkinson's Disease (PD) where they feel like they are glued to the floor. Accurate and continuous detection is needed for effective cueing to prevent or shorten FOG episodes. A convolutional neural network (CNN) was developed to detect FOG episodes in data recorded from an inertial measurement unit (IMU) on a PD patient's ankle under semi-free living conditions. Data were split into two sets: one with all movements and another with walking and turning activities relevant to FOG detection. The CNN model was evaluated using five-fold cross-validation (5Fold-CV), leave-one-subject-out cross-validation (LOSO-CV), and performance metrics such as accuracy, sensitivity, precision, F1-score, and AUROC; Data from 24 PD participants were collected, excluding three with no FOG episodes. For walking and turning activities, the CNN model achieved AUROC = 0.9596 for 5Fold-CV and AUROC = 0.9275 for LOSO-CV. When all activities were included, AUROC dropped to 0.8888 for 5Fold-CV and 0.9017 for LOSO-CV; the model effectively detected FOG in relevant movement scenarios but struggled with distinguishing FOG from other inactive states like sitting and standing in semi-free-living environments.