Privacy-Preserving Federated Learning for Cross-Dataset EEG-Based Detection of Parkinson's Disease

Research overview

The text below is the source article’s original abstract under a verified reuse license. cc by

Abstract Early and accurate detection of Parkinson’s disease (PD) is critical for timely clinical intervention. Electroencephalogram (EEG) signals serve as promising non-invasive biomarkers for PD; however, challenges such as data privacy, limited sample sizes, and inter-institutional heterogeneity in signal acquisition remain significant barriers to robust model development. To address these issues, we propose a federated learning (FL) framework that enables decentralized EEG classification, facilitating collaborative training without exchanging sensitive raw data. EEG recordings from three publicly available datasets were transformed into three-channel time–frequency representations (TFRs) via the continuous wavelet transform (CWT) and processed using a ResNet18 backbone augmented with squeeze-and-excitation (SE) modules to enhance spatiotemporal feature extraction. Model performance was rigorously validated using subject-level Monte Carlo cross-validation (MCCV) with five independent random splits. The FL framework consistently outperformed conventional centralized training across all datasets. Under the frontal–central–occipital electrode configuration, the model achieved up to 96.67% accuracy, 93.33% sensitivity, 100.00% specificity, 100.00% precision, and 96.00% F1-score at the subject level. These findings demonstrate that the proposed framework provides an effective, scalable, and privacy-preserving solution for PD diagnostics, significantly advancing the potential for robust, multi-center clinical deployment.

Related topics