A Hybrid Neuro-Symbolic Multi-Task Deep Learning Model for Parkinson's Disease Detection and Severity Classification Using Gait Signals

Parkinson’s disease (PD) is the second most prevalent neurodegenerative disorder worldwide; therefore, it is necessary to establish effective and rapid diagnostic tools to improve patient outcomes and quality of life. The current study aims to formulate an efficient deep-learning algorithm for the automated detection of PD using vertical ground-reaction-force gait signals. This study proposes a neuro-symbolic multi-class attention capsule network design for the PD detection and severity classification based on so-called vertical ground-reaction-force signals. However, prominent and discriminative gait characteristics are obtained using a comprehensive multi-domain framework that uses continuous wavelet transform-based time frequency analyses, statistical descriptors, entropy features and biomechanical gait features. The proposed architecture merges multi-scale convolutional neural networks, self-attention mechanism, bidirectional gated recurrent unit and capsule networks with an additional neuro-symbolic module to boost interpretability. A multi-task learning framework is used for the PD detection and severity classification to be performed simultaneously. The proposed model based on five-fold cross validation achieves an accuracy of 89%, precision of 89%, recall of 91%, and F1-score of 91% along with severity accuracy of 91%, which validate the effectiveness of the proposed approach to present a reliable non-invasive solution for PD diagnosis and clinical assessment.