Freezing of Gait Detection and Prediction Based on a Dilated Temporal Encoder Approach

Freezing of Gait (FoG) is a debilitating symptom of Parkinson's Disease, severely impairing mobility and increasing the risk of falls. We propose a novel framework for FoG detection and prediction using wearable inertial measurement unit (IMU) sensors, collected from 40 Parkinson's patients. Our approach leverages a Dilated Temporal Encoder (DTE) that integrates Bidirectional Dilated Temporal Convolutional Networks (Bi-DTCN), attention mechanisms, and an auxiliary loss targeting the time segment preceding a FoG (Pre-FoG). Using patient-dependent (PD) training and test sets, we systematically analyzed the effect of varying Pre-FoG segment lengths (2-5 seconds), which allowed us to identify 5 seconds as the optimal window, with a robust F1 score of 91.33 ± 0.98%. This anticipatory window has practical clinical implications, enabling timely interventions such as cueing or stopping before freezing occurs. Conversely, under patient-independent (PI) settings, experiments highlighted challenges in generalization, with performance peaking at an F1 score of 62.90 ± 1.88%. In order to address this, we introduced a performance-based clustering strategy that identified subgroups of patients with shared gait characteristics, improving cross-patient performance from 52.24 ± 2.37% to 59.62 ± 1.92%. These findings highlight the potential of combining individualized modeling and subgroup generalization to support proactive FoG management and improve patient safety.