Abstract

Parkinson’s disease (PD) is an incurable neurological disease that affects patients through motor and non-motor disturbances that may be challenging to diagnose effectively. The current conventional diagnostic system relies on clinical examination which introduces subjectivity and cannot capture small variations in the temporal dimension. Current machine learning/deep learning models have been able to diagnose Parkinson’s disease automatically through gait, speech, facial expressions, handwriting, and EEG signals. These models, however, emphasize the importance of either the spatial/local patterns or the long-term temporal dynamics but give little consideration to the multi-scale temporal patterns which may limit their robustness in the presence of disease-related patterns at multiple scales. In this paper, we propose a new model for Parkinson’s disease diagnosis based on a Multi-Scale Temporal Attention Transformer (MSTAT). The method uses parallel temporal convolutional streams for obtaining short-, medium- and long-term representations, followed by hierarchical self-attention for modeling temporal dependencies among different distributed features. A gating function dynamically weights each scale while a residual fusion layer aggregates complementary representations before classification. The subject-wise normalization and stratified evaluation are employed to avoid any information leakage. We report 97.2% accuracy, 96.7% precision, 96.4% recall, 96.5% F1-score, and 97.3% specificity in the final evaluation. Compared with CNN-GRU, LSTM, and CNN-Transformer, the proposed MSTAT achieves 4.6%, 4.3%, and 2.6% higher accuracy, respectively. Also, MSTAT increases the recall rate with respect to the three baselines by 4.8%, 4.3%, and 2.9%. This indicates the ability of MSTAT to identify Parkinson’s disease accurately. The highest specificity of 97.3% confirms the effective discrimination between the healthy people and Parkinson’s disease patients.

Authors

Thamari Thankam¹, K. Hakkins Raj²
Cihan University-Erbil, Erbil, Iraq¹, Jimma Institute of Technology, Ethiopia²

Keywords

Parkinson’s Disease, Temporal Transformer, Multi-Scale Learning, Deep Learning, Biomedical Signal Classification