Prediction of freezing of gait based on self-supervised pretraining via contrastive learning

Freezing of gait (FoG) is a kind of serious motor symptom happened to most patients with Parkinson's disease (PD). Predicting the onset of FoG is of great value since it can enable the delivery of external cues to prevent the FoG event. Most previous methods rely on hand-crafted features to predict the occurrence of FoG events. Despite the merits of high efficiency and satisfactory accuracy for these methods, the generalization of these models is problematic. To this end, this study proposed a deep-learning-based framework for predicting FoG. In this framework, the latent gait representations are first learned by self-supervised contrastive learning, and then further tuned with the labeled instances. Different augmented views of unlabeled segments of acceleration data are utilized to learn invariant gait representations in a self-supervised manner. Two different contrasting modules are designed to enforce the similarity of the gait representations between different augmented views of the same segments. The prediction of FoG is implemented by classifying the data segment into one of the three classes: 'pre-FoG', 'Walk', and 'FoG'. Our framework for FoG prediction is evaluated on the famous DAPHnet dataset and another multi-modal FoG dataset collected at Beijing Xuanwu Hospital, China (BXHC), the best overall classification performance in terms of specificity, sensitivity, and F1 score is (94.74 %, 84.61 %, and 86.19 %) on DAPHnet, and (93.08 %, 80.00 %, and 81.07 %) on BXHC when evaluated in a subject-dependent scheme. These promising results indicate the effectiveness and potential of our framework in providing an accurate and robust performance of FoG prediction.

Automated summary

This study proposes a deep-learning-based framework for predicting freezing of gait (FoG) in patients with Parkinson's disease. The framework utilizes self-supervised contrastive learning to learn latent gait representations and achieves accurate and robust FoG prediction.