A Hybrid Deep Learning Approach for Freezing of Gait Prediction in Patients with Parkinson's Disease

Abstract

The main objective of this work is to enhance the prediction of the Freezing of Gait (FoG) episodes for patients with Parkinson's Disease (PD). Thus, this paper proposes a hybrid deep learning approach that considers FoG prediction as an unsupervised multiclass classification problem with 3 classes: namely, normal walking, pre-FoG, and FoG events. The proposed hybrid approach Deep Conv-LSTM is based on the use of Convolutional Neural Network layers (CNN) and Long ShortTerm Memory (LSTM) units with spectrogram images generated based on angular axes features instead of the normal principleaxes features as the model input. Experimental results showed that the proposed approach achieved an average accuracy of 94.55% for FoG episodes early detection using Daphnet and Opportunity publicly available benchmark datasets. Furthermore, the proposed approach achieved an accuracy of 93.5% for FoG events prediction using the Daphnet dataset with the subject independent mode. Thus, the significance of this study is to investigate and validate the impact of using hybrid deep learning method for improving FoG episodes prediction.