Decoding Parkinson's Disease from EEG Signals: A Systematic Review of Machine Learning and Deep Learning Frameworks

Abstract

Parkinson's disease (PD) is the second most common neurodegenerative disorder worldwide, yet it continues to be diagnosed primarily through subjective clinical evaluation - a process that is slow, inconsistent across observers, and largely blind to early-stage disease. This paper reviews some studies published between 2021 and 2026, drawn from IEEE Xplore, IOP Publishing, Elsevier, and Springer, all of which examine automated PD diagnosis using EEG signals combined with machine learning (ML) and deep learning (DL). Across these studies, a regular methodological pattern emerges: Researchers convert raw 1D EEG time-series into 2D time-frequency representations (spectrograms) using transforms such as Gabor, Morlet Wavelet, Wavelet Scattering (WST), Smoothed Pseudo-Wigner-Ville Distribution (SPWVD), and the S-transform. Then classify the resulting images using hybrid deep learning models. Hybrid architectures particularly CNN-Transformer pipelines and Densely Linked Bidirectional LSTM (DLBLSTM) networks consistently achieve accuracies above 95%, sometimes reaching 99.97%. They outperformed standalone CNNs by roughly 4-6% and traditional ML methods by up to 9%. A key finding of this review is that Explainable AI (XAI), when combined with automated channel selection, can reduce the number of EEG electrodes by 50% with almost no loss in diagnostic accuracy, which has significant implications for portable, wearable screening devices. That said, three recurring challenges continue to limit clinical translation: datasets remain small and often homogeneous; the volume conduction effect distorts functional connectivity measurements; and the opaqueness of deep learning models undermines clinician trust. We further examined an emerging body of Graph Neural Network studies which addressed the connectivity-modelling and volume-conduction challenges directly. We concluded this review by identifying Graph Neural Networks for brain-connectivity modeling, multimodal data fusion, and XAI-driven saliency pipelines as the most promising directions for achieving clinically trustworthy, individualized PD diagnostics.