Parkinson's disease classification using optimized attention-based deep learning from EEG signals with interpretable sub-band topography.

Parkinson's disease (PD) diagnosis remains challenging because subtle neural alterations may be difficult to capture using conventional clinical assessment alone. This study proposes an attention-based deep learning framework for classifying PD from resting-state EEG with minimal preprocessing and leakage-safe evaluation. Raw EEG recordings were first partitioned at the subject level. Within each fold, the selected motor-related EEG channel was decomposed into canonical sub-bands using discrete wavelet transform, and the resulting sub-band signals were then segmented into overlapping temporal windows. Each sub-band window was transformed into a time-frequency spectrogram using the short-time Fourier transform and classified using a ResNet-101 backbone enhanced with dual channel-spatial attention. Hyperparameters were optimized using an Enhanced Adaptive Hybrid Covariance Matrix Adaptation Evolution Strategy (AH-CMA-ES), applied only within training/internal-validation subjects in each fold. Model performance was evaluated on two independent public EEG datasets, UC San Diego and University of Iowa, using subject-wise nested leave-one-subject-out cross-validation. In each fold, the held-out subject was excluded from training, augmentation, hyperparameter optimization, early stopping, and model selection. The proposed framework achieved 95.2% segment-level and 96.77% subject-level accuracy on UCSD, and 92.1% segment-level and 92.86% subject-level accuracy on Iowa, with subject-level decisions obtained by majority voting over non-augmented test segments. In addition to classification, sub-band topographical analysis provided exploratory neurophysiological interpretation across canonical EEG rhythms, revealing patterns consistent with reported PD-related oscillatory alterations. These findings suggest that resting-state EEG combined with attention-based deep learning can support robust, interpretable PD classification, while larger heterogeneous cohorts are needed to further validate clinical generalizability.