Freezing of gait detection: The effect of sensor type, position, activities, datasets, and machine learning model

• Published in: Journal of Parkinson's disease Vol. 15; no. 1; pp. 163 - 181
• Main Authors: Borzì, Luigi, Demrozi, Florenc, Bacchin, Ruggero Angelo, Turetta, Cristian, Sigcha, Luis, Rinaldi, Domiziana, Fazzina, Giuliana, Balestro, Giulio, Picelli, Alessandro, Pravadelli, Graziano, Olmo, Gabriella, Tamburin, Stefano, Lopiano, Leonardo, Artusi, Carlo Alberto
• Format: Journal Article
• Language: English
• Published: London, England SAGE Publications 01.02.2025; Sage Publications Ltd
• Subjects: Accelerometry - instrumentation; Aged; Algorithms; Ankle; Datasets; Female; Gait; Gait Disorders, Neurologic - diagnosis; Gait Disorders, Neurologic - etiology; Gait Disorders, Neurologic - physiopathology; Humans; Learning algorithms; Machine Learning; Male; Middle Aged; Monitoring, Ambulatory - instrumentation; Monitoring, Ambulatory - methods; Movement disorders; Neurodegenerative diseases; Parkinson Disease - complications; Parkinson Disease - diagnosis; Parkinson Disease - physiopathology; Parkinson's disease; Sensors; Wearable Electronic Devices; deep learning; detection; wearable sensor; freezing of gait; machine learning; Parkinson’s disease
• ISSN: 1877-7171; 1877-718X
• DOI: 10.1177/1877718X241302766

Background Freezing of gait (FoG) is a complex, frequent, and disabling motor symptom of Parkinson’s disease (PD). Wearable technology has the potential to improve FoG assessment by providing objective, quantitative, and continuous monitoring. Objective This study aims to develop a robust FoG detection algorithm that can be embedded in a simple and unobtrusive wearable sensor system and can lead to a reliable unsupervised home assessment. Methods Twenty-two subjects with PD and FoG were enrolled, equipped with four inertial modules on the ankles, back, and wrist, and asked to perform different tasks. Feature-driven and data-driven machine learning approaches were implemented, optimized, and evaluated. Further testing was conducted on two external datasets including a total of 545 FoG episodes. Results Sixteen subjects experienced FoG, providing a total number of 101 FoG events. Results demonstrated that a single sensor on the ankle, with an adequate algorithm of data analysis based on machine learning, can provide a non-invasive approach for accurate FoG detection. The model proved robust on the independent datasets, with 88–95% FoG episodes correctly detected. Interestingly, while FoG can be easily discriminated from walking, static positions, and postural transitions, turning represents a significant challenge. The high number of false alarms still represents the main limitation of the FoG recognition algorithms. Conclusions The collected dataset includes data from different sensors at different body positions. This, together with detailed labeling of tasks, activities, FoG episodes and their severity, can be a significant contribution to research on automatic FoG detection and characterization.