Advancing Parkinson's Disease Management Through Multi-Shared-Task Self-Supervised Signal Processing

This research pioneers advancing biomedical signal processing for Parkinson's Disease (PD) management through a novel Multi-shared-task Self-supervised Convolutional Neural Network-Long Short-Term Memory (CNN-LSTM) framework. This methodology leverages motion signals from wearable sen-sors, utilizing raw data and their spectrogram representations to estimate PD tremor scores continuously. By integrating unsuper-vised pre-training with dual-channel CNN, optimized for 1D and 2D signal processing, and subsequent LSTM units for temporal analysis, our approach captures complex patterns in signal data for accurate tremor assessment. The dataset comprises 526 minutes of motion data from 24 PD patients performing daily activities, processed using bandpass filtering and segmented into 5-second windows to analyze tremor-specific spectral features. Our proposed model demonstrates r=0.93 correlation between estimated and clinically assessed tremor scores, significantly outperforming traditional supervised models. This framework sets a new standard in PD tremor assessment and exemplifies the potential of advanced signal processing techniques in enhancing biomedical applications.