Freezing of Gait Detection Using Gramian Angular Fields and Federated Learning from Wearable Sensors

Area of Science:

• Biomedical Engineering
• Neurology
• Machine Learning

Background:

• Freezing of gait (FOG) is a major Parkinson's disease symptom impacting mobility and fall risk.
• Current FOG detection methods struggle with patient variability, sensor limitations, and real-world applicability.
• Existing systems often require multiple sensors, increasing failure points and limiting practical use.

Purpose of the Study:

• To develop and evaluate FOGSense, a single-sensor system for accurate, real-time FOG detection in free-living Parkinson's disease patients.
• To address limitations of existing FOG detection systems, including privacy concerns and adaptability.
• To enable continuous, in-home monitoring for improved FOG management.

Main Methods:

• Utilized Gramian Angular Field (GAF) transformations for gait pattern analysis.
• Implemented privacy-preserving federated deep learning for adaptive, collaborative model training.
• Evaluated the system on a public Parkinson's dataset collected in a free-living environment using a single sensor.

Main Results:

• FOGSense demonstrated a 22.2% improvement in F1-score and a 74.53% reduction in false positive rate compared to state-of-the-art methods.
• Achieved 10.4% higher accuracy than a single-axis accelerometer and showed robustness to missing data.
• The federated approach enabled personalized model adaptation and efficient smartphone synchronization.

Conclusions:

• FOGSense is a robust, real-world deployable system for FOG detection in Parkinson's disease.
• The single-sensor, federated learning approach enhances accuracy, privacy, and adaptability for long-term monitoring.
• This system empowers preventive care and better symptom management for individuals with Parkinson's disease.