Can explainable knowledge-driven feature pipelines match deep learning for EEG classification?

A feature pipeline built to compete across tasks

KnowEEG asks whether a deliberately knowledge-driven pipeline can stay competitive with deep learning across several EEG tasks rather than one. EEG records cortical activity non-invasively at high temporal resolution, but the signal has a low signal-to-noise ratio and varies substantially between subjects, which is why classification is hard. Deep models learn directly from raw or time-frequency data but remain opaque; earlier feature-based methods were interpretable but often performed poorly when their hand-picked features did not suit the task.

The pipeline addresses that weakness by widening the feature space instead of narrowing it. It extracts 783 per-electrode statistics drawn from time-series and EEG literature, filters them, and combines them with between-electrode connectivity statistics. A modified Random Forest then grows its trees using a balanced sampling strategy that weighs the high-dimensional per-electrode features against the connectivity features, so neither group dominates simply because it is larger. This fixed configuration is what the authors call the KnowEEG pipeline.

Five tasks, seven baselines, ten metrics

The comparison spans emotion detection (DREAMER), eyes open versus closed detection (Crowdsourced), mental workload classification (STEW), abnormal EEG classification (TUAB), and event detection (TUEV). KnowEEG is benchmarked against seven deep learning baselines, including BENDR, MAEEG, BIOT, EEGConformer, and EEG2Rep in default and pre-trained forms, plus the feature-based Catch22. Results are reported as accuracy and AUROC, with balanced accuracy and weighted F1 used where class balance makes plain accuracy misleading.

Across the ten metrics, KnowEEG is best on 5, second on 4, and outside the top two on only 1. A Wilcoxon Signed-Rank test found that neither KnowEEG nor pre-trained EEG2Rep significantly outperforms the other across the five datasets. That framing matters: the claim is parity with the strongest deep model, not uniform superiority over every baseline.

Large gains on workload, a narrow gap on abnormal EEG

On STEW mental workload classification, KnowEEG markedly outperforms all other models on both accuracy and AUROC, with standard deviations under 1 for both metrics, indicating consistent performance across seeds. It reaches 77.86 accuracy and 85.96 AUROC there. On DREAMER it leads on accuracy at 61.53 and on AUROC at 55.29, though performance varies more across seeds and sits closer to other models.

The eyes open/closed task shows a split worth noting. KnowEEG's accuracy of 92.81 trails pre-trained EEG2Rep's 94.13, but its AUROC of 96.72 is the best reported, ahead of Catch22's 95.82. On TUEV, an extremely unbalanced six-class event dataset, KnowEEG is third in balanced accuracy at 45.06 yet leads weighted F1 by a substantial margin at 78.28; the authors argue weighted F1 is the more appropriate summary there because balanced accuracy lets small classes dominate. On TUAB, KnowEEG is second to EEG2Rep on both accuracy (80.18 versus 80.52) and AUROC (87.71 versus 88.43), with the top three models separated by only 1.31% accuracy and 1.01 AUROC, which the authors treat as comparable.

Diverse features carry performance across domains

The contrast with Catch22 clarifies the mechanism the authors credit. Catch22 achieves 95.82 AUROC on the eyes open/closed task, second best, but performs poorly on all other datasets, which the paper reads as evidence that a feature-based method struggles across diverse EEG tasks when its feature set is not diverse enough. KnowEEG's breadth of per-electrode statistics plus connectivity is presented as the reason it holds up across five domains rather than one.

The computational profile is part of the argument, not an afterthought. KnowEEG needs no GPU for training, avoids the intensive pre-training typical of self-supervised models, and does not carry the quadratic O(L²) complexity of transformer architectures during iterative training. Once the deterministic feature calculation is complete, memory and training requirements remain independent of sequence length. These are stated advantages of the design; the paper does not claim the pipeline is faster in wall-clock terms for feature extraction itself.

Occipital features align with known neuroscience

Explainability here is direct rather than post-hoc: the Random Forest exposes feature importance rankings over features that already have meaning in EEG or time-series terms. The authors test whether those rankings are trustworthy on the eyes open/closed task, chosen because the two states are well studied and because KnowEEG performs well there, since explanations from poor predictions would offer little insight.

For that dataset, Functional Power Connectivity was selected as the connectivity measure during hyperparameter selection. It is computed between each electrode pair over six frequency bands (alpha, theta, delta, sigma, beta, gamma) and includes relative band power per electrode, so it describes both band power activity and functional connectivity. Analysis of the top features found occipital-region statistics to be most discriminative: 5 of the top 5, 7 of the top 10, 13 of the top 20, and 32 of the top 100 features came from the occipital region. Three of the highest-importance features, and 6 of the top 10, were permutation entropy from occipital electrodes.

The rank-1 feature, permutation entropy on channel O1 computed with dimension 6 and tau 1, had a narrower distribution for eyes open (standard deviation 0.16) than eyes closed (0.47), and a two-sample Kolmogorov–Smirnov test found the distributions differ significantly (p<0.0001). The paper also reports increased alpha synchronisation, elevated alpha-band power, and lower gamma activity during eyes closed versus eyes open, findings it says are confirmed by existing neuroscience literature. This is the paper's evidence that the pipeline's internal logic can be checked against ground truth.

Importance rankings need grouping to stay reliable

The authors are explicit that explanations are only useful when the model performs well; if performance is poor, feature importances are unlikely to reveal much about the classes. They also flag that tree-based importance measures such as the Gini Index are biased toward high-cardinality features with more split points and can suffer correlation dilution, where importance is split among highly correlated features.

Their mitigation is to analyze importances at the level of feature clusters and brain regions, such as occipital electrodes or band powers in relation to connectivity, rather than individual features. That grouped analysis is presented as more robust to permutations in single-feature rankings, and the alignment with neuroscience literature is offered as verification. Model-agnostic tools such as SHAP or LIME are suggested as additional checks.

Design choices impose further boundaries. Univariate feature filtering manages the high-dimensional space but risks discarding multivariate interactions, and the connectivity metric is selected separately from the statistical features, so interactions between the two groups are not considered during filtering. Only a single connectivity metric is used for interpretability, and selecting it via a validation set carries a theoretical bias risk that the authors say subject-wise splitting mitigates. Preprocessing choices such as window lengths, segment durations, and frequency band definitions follow established protocols and the EEG2Rep framework for fair benchmarking, but sensitivity to these parameters has not been deeply explored.

Clinical biomarkers remain the open test

The paper's own next step is to test whether KnowEEG's performance holds on additional EEG datasets, including brain-computer interface data, which differs from the five tasks studied. It also sets out a clinical program: classify neurodegenerative disease, specifically Alzheimer's and Parkinson's, partition data strictly by subject, extend evaluation beyond binary diagnosis to regression of clinical severity scales such as MDS-UPDRS, and treat high-ranking features as candidate biomarkers to be checked against established pathophysiology and then validated on independent datasets.

The authors also propose fusing pre-calculated features with latent features from a deep network, and testing whether allowing multiple connectivity metrics at once improves performance and whether any gain is task-specific. For readers weighing the current result, the practical priority is the comparison that already holds: across these five tasks, KnowEEG is a GPU-free, interpretable alternative whose performance is statistically indistinguishable from pre-trained EEG2Rep overall, with its clearest advantages on mental workload and weighted F1 for unbalanced event detection, and its explanations most credible when read at the level of brain regions rather than single features.

Sources used in this answer