No single BCI decoder dominates: per-subject tuning adds 7 accuracy points

A new arXiv study by Vasques, Barbaste, and Oullier systematically evaluates the common claim that a single EEG motor-imagery decoding pipeline is broadly superior. Using the Mother of All BCI Benchmarks (MOABB) framework, they tested 1,056 different combinations of feature extractors, scalers, and classifiers across 165 participants from three public datasets (PhysionetMI, Cho2017, Zhou2016) in the easiest evaluation regime (within-session, within-subject). They applied rigorous multi-classifier statistics—Friedman omnibus tests, Nemenyi critical-difference analysis, and Wilcoxon signed-rank tests with effect sizes—to determine if any pipeline truly stands out.

The results upend the search for a universal decoder. Covariance tangent-space projection (cov-tgsp) and Common Spatial Patterns (CSP) are the strongest families, but their ranking is dataset-dependent; on the largest cohort (PhysionetMI, 109 participants) they are statistically indistinguishable (Nemenyi p=0.27). More importantly, the single best pipeline is optimal for only 35% of participants. Nonlinear descriptors win for roughly one-third of subjects, and simply matching the pipeline to the individual yields an average accuracy gain of seven percentage points over the best fixed choice. The authors conclude that classifier and scaler choices are secondary to feature representation, and that participant-aware model selection is essential. This work provides a quantitative lower bound on the personalization problem in BCI—no one-size-fits-all decoder exists, even under ideal conditions.