Improved Deep Learning for Parkinson’s Diagnosis Based on Wearable Sensors

Abstract

Parkinson’s disease is a neurodegenerative disease that seriously affects the quality of life of patients. In this study, we propose a new Parkinson’s diagnosis method using deep learning techniques. The method takes multi-channel sensor signals as inputs, and the full convolutional and LSTM blocks of the model perceive the same time-series inputs from two different views, and connect the extracted spatial features with temporal features. In order to improve the detection performance, a channel attention mechanism was incorporated into the model, and a data augmentation approach was used to eliminate the effect of unbalanced datasets on model training. The pd vs. hc and pd vs. dd classification tasks were performed, which improved accuracy by 4.25% and 8.03%, respectively, compared to the previous best results. Both improvements were higher than the previous methods using machine learning combined with feature extraction. To utilize the available data resources more effectively, this study conducted the pd vs. hc vs. dd triple classification task for the first time, which improved the model’s ability to identify disease features. In that task, the accuracy rate reached 78.23%. The experimental results fully demonstrated the effectiveness of the proposed deep learning method for Parkinson’s diagnosis.