EEG Deep Learning Framework Classifies Parkinson’s Disease with Interpretable Analysis

Published in the journal Brain Informatics, the study introduces a method that combines optimized deep learning with interpretable sub-band topography to address the challenges of early and accurate diagnosis. The framework achieved high accuracy on two independent public datasets, demonstrating robust performance under strict subject-wise validation protocols that minimize the risk of data leakage.

Parkinson’s disease is a progressive neurodegenerative disorder affecting over 10 million people worldwide, characterized by motor symptoms such as bradykinesia, tremors, and rigidity, alongside non-motor issues like cognitive decline. Early detection remains difficult because initial symptoms are subtle and varied, often leading to delays in diagnosis. Standard clinical assessments and neuroimaging methods, while useful, can lack sensitivity in early stages and may be expensive or inaccessible to large populations. This has driven growing interest in electroencephalography, a non-invasive, low-cost technique with excellent temporal resolution for capturing real-time brain activity.

The proposed framework utilizes a ResNet-101 backbone enhanced with dual channel and spatial attention mechanisms to process EEG data. To handle the non-stationary nature of EEG signals, the researchers employed a windowing strategy that segments the signal into temporally localized intervals. This approach allows the model to extract stable patterns over time and improves robustness to transient fluctuations. Unlike previous models that rely on fixed hyperparameters or extensive preprocessing, this framework adopts a unified end-to-end approach that processes minimally preprocessed EEG data directly.

A critical aspect of the study is its evaluation protocol, which uses subject-wise nested leave-one-subject-out cross-validation to ensure rigorous testing. In this design, all recordings from a specific subject are kept within the same fold, preventing information leakage between training and testing sets. The held-out test subject is excluded from training, data augmentation, hyperparameter optimization, and model selection. This strict separation addresses the common issue in EEG studies where overlapping temporal windows can artificially inflate performance metrics.

The model was evaluated on two independent resting-state EEG datasets: the UC San Diego dataset and the University of Iowa dataset. The UCSD dataset includes 15 Parkinson’s patients and 16 healthy controls, with recordings captured in both medication-off and medication-on states. The Iowa dataset comprises 14 patients and 14 controls, recorded with a higher spatial resolution of 64 electrodes. Using these diverse datasets allows for a more rigorous assessment of the framework’s robustness across different hardware, electrode montages, and recording conditions.

Performance results indicate that the framework achieved 95.2% segment-level and 96.77% subject-level accuracy on the UCSD dataset. On the Iowa dataset, the model achieved 92.1% segment-level and 92.86% subject-level accuracy. Subject-level decisions were obtained by majority voting over non-augmented test segments, which helps reduce the impact of noisy or ambiguous individual windows. These results suggest that the model can effectively distinguish between Parkinson’s patients and healthy controls even when tested on completely unseen subjects.

Hyperparameter optimization was performed using an Enhanced Adaptive Hybrid Covariance Matrix Adaptation Evolution Strategy, or AH-CMA-ES. This method combines the global search capabilities of evolutionary strategies with the local refinement of the Adam optimizer. The optimization process was applied only within the training and internal-validation subjects of each cross-validation fold, ensuring that the test subjects did not influence the tuning process. This approach helped mitigate overfitting and accelerated convergence, allowing the network to perform consistently across different datasets.

Beyond classification, the study included a sub-band topographical analysis to provide exploratory neurophysiological interpretation. This post-hoc analysis examined the spatial distribution of power across canonical EEG frequency bands, including delta, theta, alpha, beta, and gamma. The topographical maps revealed patterns consistent with reported Parkinson’s-related oscillatory alterations, such as increased beta activity over sensorimotor regions and reduced alpha power in posterior areas. These visualizations offer qualitative physiological context for the classification findings, linking model behavior to known neurophysiological changes.

Statistical comparisons between Parkinson’s and healthy control groups further supported the topographical observations. Using Mann-Whitney U tests with Benjamini-Hochberg false discovery rate correction, the researchers found significant group-level differences in posterior alpha, central sensorimotor beta, and frontal theta activity. Parkinson’s subjects showed lower alpha power in parietal and occipital regions and higher beta power in central and frontal regions. These findings provide quantitative support for the visual patterns observed in the topographical maps, although they should be interpreted as exploratory physiological evidence rather than validated clinical biomarkers.

The study acknowledges several limitations, primarily the relatively small subject cohorts in the datasets used. Both the UCSD and Iowa datasets were collected under controlled research conditions and provide limited information regarding comorbidities, disease severity, and demographic heterogeneity. The researchers emphasize that the reported performance should be interpreted as a conservative subject-independent estimate within these specific cohorts. Further validation on larger, multicenter, and clinically heterogeneous EEG datasets is necessary to assess the framework’s broad clinical generalizability.

Despite these limitations, the study demonstrates the potential of resting-state EEG combined with attention-based deep learning for Parkinson’s disease classification. The integration of adaptive hyperparameter optimization and interpretable sub-band analysis offers a balanced approach that balances diagnostic accuracy with physiological insight. The framework’s ability to operate on minimally preprocessed data makes it a promising candidate for practical clinical screening, provided that future studies can validate its performance in more diverse and larger populations.

Future directions for this research include longitudinal EEG studies to evaluate sensitivity to disease progression and response to dopaminergic therapy. The integration of multimodal data, such as gait analysis, speech biomarkers, and wearable sensor streams, could enable a more holistic characterization of Parkinson’s disease. Prospective multicenter studies involving larger and more heterogeneous populations will be essential to determine the framework’s robustness across different clinical settings and acquisition systems.

The classification pipeline relied on a representative motor-related EEG channel, such as Cz or Pz, to evaluate the feasibility of a low-channel screening framework. This single-channel setting was intentionally adopted to assess whether a clinically simpler acquisition configuration could retain discriminative information relevant to Parkinson’s disease classification. The choice reflects a practical design trade-off between acquisition simplicity and spatial coverage, rather than a suggestion that other electrodes are uninformative. A formal comparison between single-channel, region-specific, and full-montage classification remains an important direction for future work, as the current study did not perform a direct ablation of these input configurations.

Preprocessing involved decomposing the selected EEG signal into constituent frequency bands using Discrete Wavelet Transform, generating canonical sub-band signals for delta, theta, alpha, beta, and gamma rhythms. These sub-band signals were then segmented into overlapping temporal windows and transformed into spectrograms using the Short-Time Fourier Transform. This dual transformation served complementary roles: the wavelet transform isolated physiologically meaningful EEG rhythms, while the Fourier transform generated time–frequency image representations suitable for deep learning classification. The windowing strategy addressed the intrinsic non-stationarity of EEG signals by capturing dynamic features and local bursts of activity that might otherwise be averaged out in longer recordings.

Computational analysis revealed that the primary overhead of the proposed framework was associated with the evolutionary hyperparameter optimization stage. Standard CMA-ES increased training and optimization time to approximately 58 minutes per fold on the UCSD dataset and 86 minutes per fold on the Iowa dataset. The proposed AH-CMA-ES strategy reduced these costs to 48 and 72 minutes per fold, respectively, while maintaining the best classification performance among the evaluated variants. This reduction suggests that the hybrid design, which combines evolutionary exploration with Adam-based local refinement, can decrease the number of costly search iterations compared with standard CMA-ES alone.

Despite the higher offline optimization cost, the trained model remained computationally practical for post-training subject-level prediction. Inference required only a forward pass through the trained network followed by majority voting across non-augmented test segments. The final model achieved an inference time of approximately 2.5 milliseconds per segment, with total subject-level inference times of 1.9 seconds for the UCSD dataset and 3.7 seconds for the Iowa dataset. These results indicate that the additional computational burden is confined mainly to the offline training stage, allowing for efficient deployment in settings where model development and clinical application are performed separately.

The ablation study demonstrated that the stepwise performance improvements were consistent under the subject-wise cross-validation protocol. Transitioning from a baseline CNN to ResNet-101 increased accuracy from 88.6% to 91.8%, highlighting the benefit of deep residual learning for extracting robust spectrotemporal representations. The addition of dual attention modules further improved accuracy to 93.6%, indicating a complementary effect between channel-wise recalibration and spatial feature emphasis. The final integration of AH-CMA-ES optimization raised accuracy to 95.2%, confirming that adaptive hyperparameter tuning contributed to genuine improvements in generalization to unseen subjects rather than merely reflecting segment-level leakage.

Training dynamics showed that the UCSD dataset exhibited a smoother and more stable accuracy trajectory, whereas the Iowa dataset displayed greater validation variability. This difference may reflect higher inter-subject heterogeneity, differences in acquisition conditions, and less controlled medication-state information in the Iowa cohort. The validation curves were computed using only internal validation subjects within the cross-validation protocol, ensuring that held-out test subjects were excluded from all optimization stages. The observed stability in convergence supports the effectiveness of the proposed optimization strategy in maintaining reasonable consistency between training and internal validation performance across both datasets.

Subject of Research: Deep learning classification of Parkinson's disease using resting-state EEG signals

Article Title: Parkinson’s disease classification using optimized attention-based deep learning from EEG signals with interpretable sub-band topography

Article References: Rezaee, K., Ghayoumi Zadeh, H., & Fayazi, A. (2026). Parkinson’s disease classification using optimized attention-based deep learning from EEG signals with interpretable sub-band topography. Brain Informatics, 13(1), Article 29. https://doi.org/10.1186/s40708-026-00317-x

Image Credits: AI Generated

DOI: 10.1186/s40708-026-00317-x

Keywords: Parkinson's disease, EEG, Deep learning, ResNet-101, Attention mechanism, Neurodegeneration, Signal processing, Medical imaging, Machine learning, Brain Informatics, Parkinson, disease