An Efficient FoG-M3 Method for Self-Adaptive Labeling and Predicting Freezing of Gait

Area of Science:

• Neurology
• Biomedical Engineering
• Artificial Intelligence

Background:

• Freezing of gait (FoG) is a debilitating motor symptom in Parkinson's disease (PD), increasing fall risk.
• Existing FoG identification methods struggle with accurate prediction due to labeling complexities, data imbalance, and limited feature representation.

Purpose of the Study:

• To develop an accurate and efficient deep learning-based method for predicting freezing of gait (FoG) in Parkinson's disease patients.
• To overcome challenges in Pre-FoG labeling, data imbalance, and model feature representation.

Main Methods:

• Introduced FoG-M3, a deep learning model integrating FoG-Mix, U-Net, Mamba module, and MoCo contrastive learning.
• Utilized non-fixed length for Pre-FoG labeling to improve accuracy.
• Applied contrastive learning to enhance model representational capacity.

Main Results:

• Achieved 95% overall prediction accuracy on the Daphnet dataset and 93% on the BHXC dataset.
• Reached approximately 90% accuracy in Pre-FoG prediction on both datasets.
• Demonstrated prediction of FoG events approximately 6 seconds prior to onset, enabling timely intervention.

Conclusions:

• The FoG-M3 method offers high accuracy and efficiency in predicting Pre-FoG, surpassing existing models.
• This predictive capability holds significant potential for improving Parkinson's disease management and rehabilitation.
• The approach provides a substantial time window for patients to adjust posture and prevent falls.