A Novel Spatio-Temporal Based AI Framework For Parkinson Disease Detection Using Physiological Signals

Parkinson’s Disease (PD) is a progressive neurological disorder characterized by motor and cognitive impairments. Early diagnosis is critical for timely intervention and better patient outcomes, yet it remains challenging due to subtle neurological changes in early stages. Electroencephalography (EEG) offers a promising noninvasive method for detecting PD related brain activity changes. However, its high dimensionality and non-stationary nature complicate analysis. In this thesis we have designed and developed an interpretable machine learning framework for automatic PD detection using EEG data. We developed a spatio-temporal architecture combined with various feature extraction methods such as Wavelet Packet Decomposition (WPD), Empirical Mode Decomposition (EMD), Wavelet Scattering Transform (WST), and Power Spectral Density (PSD). We have found the proposed spatio-temporal architecture (Archi-3) performed extremely well, especially with EMD and WPD, by learning from the spatial and temporal EEG characteristics specific to PD. Our proposed architecture (Arch-3) achieved 100% accuracy, precision, recall, and F1 score using WPD features for PD classification. Additionally, under subject-wise cross-validation, it outperformed all other methods using EMD features, achieving 72.78% accuracy, 70.79% precision, 86.67% recall, and 75.96% F1 score, demonstrating strong spatial and temporal modeling capabilities. For model transparency, we have used SHAP analysis which identified key EEG electrodes FC2, Fp2, P4, and C4, across models to be more influential and significant for PD. This finding also aligns with known brain regions affected by PD. Additionally, we developed a web-based application for real time prediction and showing the reason behind the models prediction to facilitate clinical use. In future research we will target hyperparameter optimization through metaheuristic algorithms, improved feature selection to reduce computational demands.

Keyword

Graph Neural Networks, Recurrent Neural Networks, Parkinson, SHAP, Wavelet Packet Decomposition, Empirical Mode Decomposition, Wavelet Scattering Transform.

• Supervision