DEEP CHANNEL AND TEMPORAL ATTENTION TO DETECT PARKINSON’S DISEASE USING GAIT SIGNALS

Abstract

Parkinson’s disease (PD) is a neurodegenerative disorder that changes constantly. A general and clinically relevant gait impairment is a significant symptom, emerging as a non-invasive method to support early diagnosis and monitoring of PD. In this paper, an enhanced deep learning (DL) framework that combines convolutional neural networks (CNNs) and bidirectional long short-term memory (BiLSTM) with dual attention (DA) fusion consisting of channel-temporal attention (CNN-BiLSTM-CTA) is implemented to classify PD versus healthy controls (HC) using the signals' vertical ground reaction force (VGRF).  The model was trained on gait recordings, in which VGRF signals were segmented into 300-sample windows with 50% overlap using basic augmentation and focal loss. Unlike prior studies that relied on pooled cross-validation, we adopted a rigorous protocol by training and validating on the Ga and Ju cohorts while holding back the Si cohort for independent testing. The results demonstrate that CNN–BiLSTM-CTA enhances subject-level performance across accuracy, sensitivity, specificity, and F1-score compared to the baseline CNN, BiLSTM, and CNN–BiLSTM models, as well as prior reported transformer-based approaches that achieve accurate gait signal-based PD identification.