Work

Nonlinear Complexity Features Add Nothing beyond the Power Spectrum in Motor-Imagery BCI: A Surrogate-Controlled Reassessment

Year: 2026
Type: article
Abstract: A significant amount of research in the field of Motor-Imagery (MI) brain–computer interfaces have shown that nonlinear and higher-order features-such as fractal dimensions, entropies, detrended fluctuation analysis, variational-mode complexity, Phase–Amplitude Coupling (PAC), and bispectrum/bicoherence-enhance the decoding of sensorimotor rhythms. However, the past studies have drawn their conclusions based on within-session cross-validation, per-subject tuning, fusion over a spatial–spectral baseline, and the absence of a surrogate control. The present study poses a more focused and testable question: do these features contain information that goes beyond that provided by the power spectrum when assessed fairly? Using the BCI Competition IV datasets 2a and 2b (18 subjects) under the official independent-session protocol, a single matched shrinkage-LDA classifier is applied to decode fifteen families of features, each family is subjected to an Iterative Amplitude-Adjusted Fourier Transform (IAAFT) surrogate test, and all the families of tests are corrected for multiple comparisons. The results are clear. Nonlinear-7 achieves 0.576 compared to 0.691 for plain band power on dataset 2a and 0.668 compared to 0.735 on dataset 2b; when compared to a filter-bank CSP and Riemannian ensemble, the additional value of the surrogate-corrected nonlinear and PAC candidates is only +0.0000 Cohen's κ, and mutual-information selection includes them in 0.0% of the feature slots even though they make up 33.3% of the candidate pool. The higher-order spectra make the mechanism evident: across all 18 subjects, the magnitude of the bispectrum is statistically no different from that of band power (0.715 versus 0.721, p=0.59) and the accuracy on a per-subject basis correlates with band-power accuracy at r=0.958; in contrast, the normalized bicoherence, which removes that power scaling, achieves a decoding accuracy of 0.546-significantly lower than band power (p=3.5×10−6)-and falls below the level expected from its spectrally matched null (z=−0.16). A montage factorial that separates electrode count from spatial derivation shows that the apparent competitiveness of the complexity features on low-channel montages is not due to spatial focusing. This indicates the robustness of the results: when the number of channels, the number of electrodes, and the spatial derivation are kept the same across datasets, the ordering between nonlinear and band power remains unchanged, and the accuracy of the nonlinear features correlates with that of band power across the entire montage ladder (r=0.77 on 2a, r=0.84 on 2b). The results do not mean that all useful features disappear after controlling for the spectral information. In 3 of the 4 dataset-and-descriptor combinations, a small amount of information remains above chance even after correcting for false discoveries. However, this leftover effect is very small-only 4 to 6 percentage points above chance. By comparison, band power is 19 to 24 points above chance. And when these remaining features are added to the best-performing pipeline, they provide no additional benefit. This paper reports that nonlinear complexity features do not carry any usable information beyond the power spectrum for MI decoding, and the authors make available a reusable surrogate redundancy test that any candidate feature should pass before a nonlinear gain is claimed.
Language: English
Open Access status: bronze
OpenAlex ID: W7226005881
FWCI:
Cites:
Cited by:
Related to: 10