Transformer-Based Classification of Parkinson’s Disease from EEG Using BIDS-Formatted OpenNeuro Datasets

Abstract

Artificial intelligence systems that support rapid and reliable decision-making under time-sensitive constraints must interpret high-dimensional sensory inputs while remaining robust to variability and noise. Electroencephalography (EEG) exemplifies such a sensing modality due to its non-stationarity, subject variability, and sensitivity to recording conditions. This study evaluates a Transformer-based framework for multi-class classification of Parkinson’s disease medication state (PD-ON, PD-OFF) and healthy controls using structured spectral and entropy-based EEG features. Rather than proposing an end-to-end raw-signal model, we assess whether attention-based sequence modeling offers advantages over classical classifiers when operating on physiologically interpretable representations. Three heterogeneous BIDS-formatted OpenNeuro datasets (PD1: n = 31; PD2: n = 50; PD3: n = 56) were evaluated using subject-wise grouped cross-validation to prevent participant-level data leakage. The Transformer achieved near-ceiling F1-scores on relatively homogeneous datasets (PD1: 99.33%; PD3: 98.35%) and competitive performance on the more heterogeneous PD2 dataset (88.36%), underscoring the impact of sensing variability on robustness. Attention analysis indicates adaptive weighting of informative temporal–spectral features. These findings provide a systematic empirical evaluation of attention-based sequence modeling under heterogeneous sensing conditions.