OBJECTIVE

Freezing of Gait (FOG) is prevalent in people with Parkinson's disease (PD) and severely disrupts mobility. Detecting the exact boundaries of FOG episodes may facilitate new technologies in "breaking" FOG in real-time. This study investigates the performance of automatic device-based FOG detection.

METHODS

Eight machine-learning classifiers (including Neural Networks, Ensemble methods, and Support Vector Machines) were developed using (i) accelerometer and (ii) combined accelerometer and gyroscope data from a waist-worn device. While wearing the device, 107 people with PD completed mobility tasks designed to elicit FOG. Two clinicians independently annotated exact FOG episodes using synchronized video and a flowchart algorithm based on international guidelines. Device-detected FOG episodes were compared to annotated episodes using 10-fold cross-validation and Interclass Correlation Coefficients (ICC) for agreement.

RESULTS

Development used 50,962 windows of data and annotated activities (>10 hours). Strong agreement between clinicians for precise FOG episodes was observed (90% sensitivity, 92% specificity, and ICC1,1 = 0.97 for total FOG duration). Device performance varied by method, complexity, and cost matrix. The Neural Network using 67 accelerometer features achieved high sensitivity to FOG (89% sensitivity, 81% specificity, and ICC1,1 = 0.83) and 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 strong clinician agreement indicates room for improvement in future device-based FOG detection.

SIGNIFICANCE

This study may enhance PD care by reducing reliance on visual FOG inspection, demonstrating that high sensitivity in automatic FOG detection is achievable.