Development, Validation, and Limits of Freezing of Gait Detection Using a Single Waist-Worn Device

Objective: Freezing of Gait (FOG) often described as the sensation of “the feet being glued to the ground” is prevalent in people with Parkinson's disease (PD) and severely disturbs mobility. In addition to tracking disease progression, precise detection of the exact boundaries for each FOG episode may enable new technologies capable of “breaking” FOG in real time. This study investigates the limits of sensitivity and performance for automatic device-based FOG detection. Methods: Eight machine-learning classifiers (including Neural Networks, Ensemble & Support Vector Machine) were developed using (i) accelerometer and (ii) accelerometer and gyroscope data from a waist-worn device. While wearing the device, 107 people with PD completed a walking and mobility task designed to elicit FOG. Two clinicians independently annotated the precise FOG episodes using synchronized video according to international guidelines, which were incorporated into a flowchart algorithm developed for this study. Device-detected FOG episodes were compared to the annotated FOG episodes using 10-fold cross-validation to determine accuracy and with Interclass Correlation Coefficients (ICC) to assess level of agreement. Results: Development used 50,962 windows of data representing over 10 hours of data and annotated activities. Very strong agreement between clinicians for precise FOG episodes was observed (90% sensitivity, 92% specificity and ICC 1,1 = 0.97 for total FOG duration). Device-based performance varied by method, complexity and cost matrix. The Neural Network that used only 67 accelerometer features provided a good balance between high sensitivity to FOG (89% sensitivity, 81% specificity and ICC 1,1 = 0.83) and solution stability (validation loss ≤ 5%). Conclusion: The waist-worn device consistently reported accurate detection of precise FOG episodes and compared well to more complex systems. The superior agreement between clinicians indicates there is room to improve future device-based FOG detection by using larger and more varied data sets. Significance: This study has clinical implications with regard to improving PD care by reducing reliance on clinical FOG assessments and time-consuming visual inspection. It shows high sensitivity to automatically detect FOG is possible.