Parkinson’s disease detection from resting-state EEG using a Transformer model with Multi-Head Attention explanations

Despite the proven potential of using Deep Learning (DL) models based on electroencephalographic (EEG) signals to detect neurological disorders like Parkinson’s Disease (PD), their adoption in clinical practice is limited due to insufficient reliability and generalizability. We propose an interpretable end-to-end DL framework leveraging the Multi-Head Attention (MHA) component of Transformers to classify EEG signals of 100 PD patients and 79 control subjects across three public resting-state EEG datasets. A systematic interpretability approach, including embedding visualization and MHA-based temporal and spectral analysis through statistical tests, is proposed to enhance the identification of discriminative biomarkers. Experimental findings across a large, multi-centric cohort of subjects demonstrated the framework’s capability to detect meaningful EEG patterns in the frequency intervals of interest. Interpretability analysis revealed that MHA focused on specific temporal patches in the input signal, which correlated to the classification outcomes. Spectral analysis identified significant power differences in Theta and Beta bands, capturing neural patterns of cognitive and motor dysfunction in PD. Furthermore, attention-guided segmentation improved the sensitivity of spectral biomarkers, such as Alpha/Theta ratio and Beta relative power, consistent with prior literature. Moreover, the proposed approach yielded the highest epoch-level mean AUCs of 0 . 91 ± 0 . 01 on Theta, 0 . 84 ± 0 . 03 on Alpha, and 0 . 81 ± 0 . 01 on All-band, achieving state-of-the-art performances while also demonstrating robustness to heterogeneous data.