A Time-Frequency Decoupled Contrastive Learning Framework for Electroencephalography-Based Parkinson's Disease Diagnosis.

Parkinson's disease (PD), a prototypical neurodegenerative disorder, poses significant challenges for early diagnosis. Motivated by recent advances in representation learning, a novel EEG-based diagnostic framework for PD, termed the time-frequency decoupling contrastive learning network (TDCNet), is introduced. The framework adopts collaborative contrastive learning between an online network and a momentum network to optimize latent representations in a self-supervised manner. It incorporates a temporal-evolution contrastive loss together with a frequency-aware contrastive loss, enabling adaptive decoupling of dynamic fluctuations and rhythmic patterns in EEG signals. Furthermore, structured pseudo-labels are constructed at the subject level to promote aggregation of intra-subject samples within the latent space. Ground-truth labels are introduced only in a downstream classifier for supervised evaluation. TDCNet achieves accuracies ranging from 95% to 99% across three public datasets under subject-dependent evaluation and 73% to 89% under subject-independent evaluation, consistently outperforming existing approaches. Additional analyses on datasets with available medication annotations (UC and UNM) suggest that medication status may potentially contribute to improved diagnostic performance. Brain-region investigations further reveal that central electrodes play a dominant role, yielding accuracies of 78% to 88% under subject-independent settings. Extensive ablation studies and comparative experiments confirm the effectiveness of the proposed framework and provide new insights into EEG-based PD diagnosis. The code is publicly available at https://github.com/Mike007-netizen/TDCNet.