A Swap-Adversarial Framework for Improving Domain Generalization in Electroencephalography-Based Parkinson's Disease Prediction

Electroencephalography (ECoG) offers a promising alternative to conventional electrocorticography (EEG) for the early prediction of Parkinson’s disease (PD), providing higher spatial resolution and a broader frequency range. However, reproducible comparisons has been limited by ethical constraints in human studies and the lack of open benchmark datasets. To address this gap, we introduce a new dataset, the first reproducible benchmark for PD prediction. It is constructed from long-term ECoG recordings of 6-hydroxydopamine (6-OHDA)-induced rat models and annotated with neural responses measured before and after electrical stimulation. In addition, we propose a Swap-Adversarial Framework (SAF) that mitigates high inter-subject variability and the high-dimensional low-sample-size (HDLSS) problem in ECoG data, while achieving robust domain generalization across ECoG and EEG-based Brain-Computer Interface (BCI) datasets. The framework integrates (1) robust preprocessing, (2) Inter-Subject Balanced Channel Swap (ISBCS) for cross-subject augmentation, and (3) domain-adversarial training to suppress subject-specific bias. ISBCS randomly swaps channels between subjects to reduce inter-subject variability, and domain-adversarial training jointly encourages the model to learn task-relevant shared features. We validated the effectiveness of the proposed method through extensive experiments under cross-subject, cross-session, and cross-dataset settings. Our method consistently outperformed all baselines across all settings, showing the most significant improvements in highly variable environments. Furthermore, the proposed method achieved superior cross-dataset performance between public EEG benchmarks, demonstrating strong generalization capability not only within ECoG but to EEG data. The new dataset and source code will be made publicly available upon publication.

💡 Research Summary

Parkinson’s disease (PD) is a progressive neurodegenerative disorder whose early detection is crucial for effective treatment. While electroencephalography (EEG) is widely used for non‑invasive PD prediction, its limited spatial resolution and attenuation of high‑frequency components hinder the capture of disease‑specific biomarkers such as β‑band fluctuations and high‑frequency oscillations. Electrocorticography (ECoG) overcomes these limitations by recording directly from the cortical surface, providing millimeter‑scale spatial resolution and access to frequencies above 200 Hz. However, human ECoG studies are constrained by ethical and safety concerns, and existing animal datasets suffer from heterogeneous channel configurations and long‑term non‑stationarity, resulting in a high‑dimensional low‑sample‑size (HDLSS) problem and severe inter‑subject variability.

To address these challenges, the authors make three major contributions. First, they construct and publicly release MOCOP, the first reproducible benchmark for PD prediction based on ECoG. The dataset comprises long‑term recordings from eleven 6‑hydroxydopamine (6‑OHDA) lesioned rats, each annotated with “pre‑stimulation” and “post‑stimulation” states. Channel layouts differ across subjects, and recordings contain artifacts and temporal drift, reflecting realistic HDLSS conditions.

Second, they introduce Inter‑Subject Balanced Channel Swap (ISBCS), a novel data‑augmentation technique. ISBCS aligns channels that correspond to the same anatomical location across subjects and randomly swaps them, generating counterfactual “virtual” subjects. This operation (a) preserves spatial relationships, (b) balances anatomical and electrode‑placement differences, and (c) forces the model to focus on disease‑related patterns rather than subject‑specific idiosyncrasies. By enriching the distribution of training samples, ISBCS mitigates over‑fitting in the HDLSS regime.

Third, they embed domain‑adversarial learning into a Swap‑Adversarial Framework (SAF). An encoder and task classifier are trained jointly on the PD prediction task, while a domain discriminator (predicting subject identity) is attached via a Gradient Reversal Layer (GRL). The adversarial objective drives the encoder to discard subject‑specific cues while retaining discriminative disease information. Prior to augmentation, the authors apply Artifact Subspace Reconstruction (ASR) to remove high‑energy low‑dimensional artifacts, ensuring that the domain discriminator does not latch onto noise.

Extensive experiments validate the approach. In leave‑one‑subject‑out cross‑subject evaluation on the ECoG dataset, SAF outperforms a range of baselines—including CNN, LSTM, MMD‑based domain alignment, and Group DRO—by 8–12 percentage points in accuracy. Cross‑session tests (wireless vs. wired acquisition) demonstrate robustness to hardware and environmental changes. Crucially, the same framework transferred to two public EEG benchmarks (BCI‑IV and PhysioNet), achieving >85 % accuracy despite modality shift, thereby confirming strong cross‑modal generalization.

In summary, the paper delivers (1) a publicly available ECoG PD benchmark, (2) a channel‑swap augmentation (ISBCS) that alleviates HDLSS and inter‑subject bias, and (3) a domain‑adversarial training scheme that learns subject‑invariant yet disease‑relevant representations. The combined Swap‑Adversarial Framework sets a new standard for domain‑generalizable brain‑signal classification and opens avenues for applying similar techniques to other modalities such as MEG or fNIRS, as well as for integrating meta‑learning or prompt‑based adaptation to further enhance clinical applicability.