TopoAdapter: a plug-and-play multi-hop topology adapter for MI-EEG decoding.

Background

Motor imagery electroencephalography (MI-EEG) decoding is limited by low signal-to-noise ratio, non-stationarity, inter-subject variability, and small calibration sets. Lightweight decoders are attractive for online BCI but often learn channel relations only from limited training data. NEW

Method

We introduce TopoAdapter, a plug-and-play input module that injects a fixed electrode-layout prior into existing EEG backbones. It builds a physical electrode graph from the montage, computes cumulative multi-hop channel-to-neighborhood contrasts, and adds a learnable low-amplitude residual while preserving input shape.

Results

Under an aligned 500-epoch, five-seed ATCNet protocol, mean accuracy changes were +1.17 and +0.13 percentage points on the BCI Competition IV-2a and Zhou2016 motor-imagery datasets, respectively, and +0.36 points on the High-Gamma executed-movement dataset. All three dataset-level means were positive; the BCI IV-2a and High-Gamma bootstrap intervals excluded zero, although no paired test remained significant after three-dataset Holm correction. In a separate BCI IV-2a compatibility study, all seven selected backbones improved on average and three retained Holm-adjusted Wilcoxon evidence. COMPARISON WITH EXISTING

Methods

Unlike graph neural decoders that redesign the backbone, TopoAdapter keeps downstream components unchanged. In a matched comparison with the open-source Adaptive Channel Mixing Layer (ACML), TopoAdapter attained 60.52% versus 60.42% accuracy with 70 versus 506 added parameters; the direct paired difference was unresolved.

Conclusions

TopoAdapter provides an explicit, ultra-lightweight spatial prior for motor EEG decoding. Positive mean changes on two motor-imagery datasets, one executed-movement dataset, and seven heterogeneous BCI IV-2a backbones support portability and a favorable cost-benefit profile, while effect magnitude remains dataset- and subject-dependent.