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

Freezing of gait (FoG) is a common motor impairment that occurs as Parkinson's disease patients enter the moderate stage or higher, leading to a high risk of falls. Although previous methods have accurately identified FoG, significant challenges remain in predicting it. These challenges include difficulties in labeling Pre-FoG due to its personalization, sudden onset, and variability; the imbalance between the amount of FoG and normal gait data; and inadequate feature representation by the model. To address these challenges, we introduce a deep learning-based method for FoG prediction, named FoG-M3 (Mixup, MoCo, and MU-Net). This method employs a non-fixed length for Pre-FoG rather than a fixed length, allowing for more accurate labeling. To address the long-tail issue in the dataset, we integrated FoG-Mix with U-Net and the Mamba module. Additionally, the FoG-MoCo contrastive learning strategy was applied to enhance the model's representational capacity. The overall prediction accuracy reaches 95% on the Daphnet dataset and 93% on the BHXC dataset. Pre-FoG prediction accuracy also achieves approximately 90% on both datasets, surpassing the performance of FoG-Net, ResNeXt, and contrastive learning-based ResNet. Using the proposed method, we can predict FoG occurrences with enough time for posture adjustments, approximately 6 seconds before onset, which is more than typically available to Parkinson's disease patients. Therefore, the FoG-M3 method can predict Pre-FoG with high accuracy and efficiency, demonstrating substantial potential for use in Parkinson's disease rehabilitation.