KnowEEG: Explainable Knowledge Driven EEG Classification

Amarpal Sahota; Navid Mohammadi Foumani; Raul Santos-Rodriguez; Zahraa S. Abdallah

cs.LGMay 1, 2025

KnowEEG: Explainable Knowledge Driven EEG Classification

Authors: Amarpal Sahota, Navid Mohammadi Foumani, Raul Santos-Rodriguez, Zahraa S. Abdallah

Organizations: University of Bristol, Bristol, United Kingdom · Monash University, Melbourne, Victoria, Australia

Abstract

Electroencephalography (EEG) classification has shown significant promise in applications ranging from neurodegenerative disease detection to brain-computer interfaces (BCIs). Whilst deep learning methods have increasingly dominated the research landscape in recent years; these models remain largely opaque 'black boxes' with limited interpretability. In this paper, we revisit the utility of feature-based classification via a systematic pipeline designed to integrate per-electrode signal statistics with cross-brain connectivity metrics. The proposed "KnowEEG" pipeline extracts 783 per-electrode statistical features alongside between-electrode connectivity statistics, utilizing a balanced sampling strategy between high-dimensional per electrode features and connectivity. This feature-based pipeline achieves performance comparable to or exceeding state-of-the-art deep learning models across five diverse classification tasks: emotion detection, mental workload classification, eyes open/closed detection, abnormal EEG classification, and event detection. This performance is achieved without the need for Graphics Processing Unit (GPU) resources. Furthermore, the pipeline provides inherent explainability through feature importance rankings. We demonstrate via the eyes open/closed task that this explainability identifies discriminative biomarkers in the occipital region that are consistent with established neuroscience literature. These results suggest that despite significant advances in deep learning, structured feature engineering remains a high performance and interpretable method for EEG classification.

Figures & tables

Figure 1: Feature Based Pipeline KnowEEG : There are two parallel threads to the pipeline. In thread one, 783 features from generalised and EEG-specific time-series literature are calculated per electrode and concatenated. Feature relevance is evaluated via univariate non-parametric hypothesis tests (Mann-Whitney U for binary targets and Kruskal-Wallis tests for multi-class targets) using only the training split of the data. The Benjamini–Yekutieli correction to control the false discovery rate is applied ( $\alpha=0.05$ ) with uninformative features filtered out. In thread two, between-electrode connectivity features are calculated for connectivity metrics. The best-performing metric is selected using classifier performance on the validation data. The features from thread one and thread two are then combined using a balanced sampling strategy (see Algorithm 1 ).

Dataset	Classification Task	Dim.	Freq.	Duration	Samples
DREAMER	Emotion Detection	14	128Hz	2s	77,910
STEW	Mental Workload	14	128Hz	2s	26,136
Crowdsourced	Eyes Open/Closed	14	128Hz	2s	12,296
TUEV	Event Detection	16	256Hz	5s	112,464
TUAB	Abnormal EEG	16	256Hz	10s	409,455

Table 1: Properties of the EEG datasets used (adapted from [ 40 ] ). DREAMER [ 29 ] , STEW [ 35 ] , Crowdsourced [ 54 ] , TUEV [ 25 ] and TUAB [ 36 ]

Table 2: Performance of the feature based KnowEEG pipeline versus competitor models across datasets. KnowEEG results are shaded in blue. Best performance per dataset is in bold with second best underlined with Accuracy results in table a) and AUROC in table b).

Figure 2: This histogram displays the Gini importance values for the Functional Power Connectivity features (FPC) used by the trained KnowEEG model on the Crowdsource dataset. The x-axis represents the feature importance score with the y-axis showing the count of features at that importance level. The distribution is highly skewed with an exponential-like decay, showing that a large majority of features have an importance score near zero, while a small group of features have higher importance scores.

Figure 3: Feature Importance score per power band for Functional Power Connectivity on the Crowdsource dataset. FPC gamma features have the highest feature importance rank score overall followed by alpha, beta, delta, theta and then sigma with the lowest score.

Figure 4: This plot shows mean relative band power in the alpha and gamma bands for eyes closed (left) and eyes open (right) participants. For the Alpha band the eyes closed group have a higher mean power particular in the Occipital Lobes towards the back of the head. For the Gamma band the eyes open group have a higher mean power with peaks in the Left and Right Temporal / Central regions of the brain.

Figure 5: This plot shows the mean Alpha and Gamma band connectivity for the eyes closed (left) and eyes open (right) groups per the FPC features (defined in B ). The eyes closed group have higher Alpha connectivity across the brain particularly in the Occipital regions. For Gamma connectivity the differences between the groups are not visually significant.

Rank	Feature	Parameters	Channel	Brain Region
1	Permutation Entropy	Dimension 6 Tau 1	O1	Occipital
2	Permutation Entropy	Dimension 4 Tau 1	O2	Occipital
3	Permutation Entropy	Dimension 5 Tau 1	O2	Occipital
4	Partial Autocorrelation	Lag 2	O1	Occipital
5	Number Peaks	Support 1	O2	Occipital
6	Permutation Entropy	Dimension 5 Tau 1	O1	Occipital

Table 3: Statistical features with highest importance ranked 1 to 10. The table displays the feature name, parameters specifying feature calculation, corresponding electrode and corresponding brain region. It is notable that the top 7 features are all from the Occipital brain region.

Figure 6: Count of features in top N importances (N = 5, 10, 20, 100) by Brain Region. From this bar chart it is clear that the Occipital region has a disproportionately large number of highly discriminative features vs. other regions. The chart shows that 5 out of the top 5 (dark blue) most important features are from the Occipital region as well as 7 of the top 10, 13 of the top 20 and 32 of the top 100.

Figure 7: KDE plot of the rank 1 feature from Table 3 : Permutation Entropy of the O1 (Occipital Lobe) EEG channel. The distribution for the ‘eyes open’ class (green) is distinctly narrower and has a higher peak density, while the ‘eyes closed’ class (blue) shows a broader distribution with a wider spread of values.

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Dataset	Signal Duration	Freq	Segment Duration	Num Segments
DREAMER	2s	128Hz	0.25	8
STEW	2s	128Hz	0.25	8
Crowdsourced	2s	128Hz	0.25	8
TUEV	5s	256Hz	0.50	10
TUAB	10s	256Hz	0.625	16

Appendix

Table 4: Overview of segmentation for signals for each dataset. Required in order to calculate specific connectivity metrics (Coherence, Imaginary Coherence, Pairwise Phase Consistency, Phase Locking Value, Phase Lag Index, Directed Phase Lag Index, Weighted Phase Lag Index).

Symbol	Definition	Context
$N$	Number of samples	Total dataset size
$K$	Number of channels	EEG electrodes (14–16)
$L$	Sequence length	Data points per sample
$M$	Total features	$A$ (per-electrode) + $B$ (connectivity)
$T$	Number of trees	Random Forest hyperparameter
$C$	Candidate metrics	Connectivity metrics for local selection

Appendix

Table 5: Definitions for Complexity Analysis

Metric	KnowEEG	EEG Conformer	EEG2Rep
Feature Extraction	$O(N\cdot K\cdot L\log L)$	$O(N\cdot K\cdot L)$	$O(N\cdot K\cdot L)$
Tuning (Search)	$O(C\cdot N\log N)$	$O(H\cdot E\cdot N\cdot L^{2}\cdot d)$	$O(H\cdot E\cdot N\cdot L^{2}\cdot d)$
Training Time	$O(T\cdot\sqrt{M}\cdot N\log N)$	$O(E\cdot N\cdot(L\cdot d+L^{2}\cdot d))$	$O(E\cdot N\cdot L^{2}\cdot d)$
Peak Memory	$O(N\cdot M)$	$O(L^{2}+\text{weights})$	$O(L^{2}+\text{weights})$
Hardware	CPU	GPU (High VRAM)	GPU (High VRAM)

Appendix

Table 6: Complexity and Memory Comparison: KnowEEG vs. Deep Learning Baselines

Performance Metric	KnowEEG	EEG Conformer	EEG2Rep
Peak Memory	3 GB (CPU RAM)	8 GB (GPU VRAM)	10 GB (GPU VRAM)
Memory Type	CPU RAM	Dedicated GPU VRAM	Dedicated GPU VRAM
Training Overhead	Exactly 1 Training Pass	E Epochs	E Epochs + Pre-training

Appendix

Table 7: KnowEEG vs. Deep Learning Baselines memory requirements on Crowdsourced Dataset

Electroencephalography

Electroencephalography (EEG) studies brain-generated electrical activity to understand cognition and support diagnosis, monitoring, and brain-computer interfaces.

See all 178 papers

Interpretable EEG biomarkers with bag-of-waves: Spatial and temporal waveform dictionaries for low-data regimes
Jul 24, 2026Athanasios Papastathopoulos-Katsaros, Steven T. Lee, Lin Yao +4Electroencephalography Interpretable ML
Department of Pediatrics, Baylor College of Medicine, Houston, TX, USA · Jan and Dan Duncan Neurological Research Institute, Texas Children’s Hospital, Houston, TX, USA · Department of Child Neurology and Developmental Neurosciences, Texas Children’s Hospital, Houston, TX, USA +3
Comparing Post-Hoc Explainable AI Methods for Interpreting Black-Box EEG Models in Depression Detection
May 27, 2026Antonia Šarčević, Nikolina FridGradient-Based Attribution Electroencephalography
University of Zagreb Faculty of Electrical Engineering and Computing, Croatia
Hierarchic-EEG2Text: Assessing EEG-To-Text Decoding across Hierarchical Abstraction Levels
Feb 24, 2026Anupam Sharma, Harish Katti, Prajwal Singh +2EEG Decoding EEG-to-Text Decoding
IIT Gandhinagar Gandhinagar, Gujarat, India · NIMH, NIH USA

Jul 24, 2026J/K move · Enter open · S save

Interpretable EEG biomarkers with bag-of-waves: Spatial and temporal waveform dictionaries for low-data regimes

Athanasios Papastathopoulos-Katsaros, Steven T. Lee, Lin Yao, Ajay Thomas+3

Department of Pediatrics, Baylor College of Medicine, Houston, TX, USA · Jan and Dan Duncan Neurological Research Institute, Texas Children’s Hospital, Houston, TX, USA · Department of Child Neurology and Developmental Neurosciences, Texas Children’s Hospital, Houston, TX, USA+3

Electroencephalography Interpretable ML Time Series Representation Learning

Open paper AskPDF arXiv

Electroencephalography (EEG) is widely used to diagnose neurological conditions, but its analysis usually relies on either predefined spectral features or deep neural networks. Predefined features carry a strong bias, since they fix in advance what counts as informative, while deep neural networks and foundation models are hard to interpret and need large amounts of data and compute. We present bag-of-waves, an interpretable framework that learns a small dictionary of recurring EEG waveform templates, called atoms, using shift-invariant k-means without labels. The continuous EEG is then turned into a sequence of atom tokens, whose counts feed a simple downstream classifier or clustering step. We extend this representation in two ways: we add atom-to-atom transitions, which we call n- grams, to capture temporal structure, and we move from single-channel atoms to regional and cross-channel spatial atoms for the multichannel case. We test the method on three complementary datasets, each probing a different aspect: single-channel mouse genotype clustering with only sixteen animals (the low-data and temporal case), resting-state dementia classification (the spatial case), and the TUEV benchmark, a six-way classification of clinical EEG events (a high-data comparison against strong deep and foundation baselines). Across all three datasets, bag-of-waves achieves performance competitive with state-of-the-art deep and foundation models. Yet, it operates with a fraction of the parameter count and provides full interpretability: because every atom corresponds to an inspectable waveform, the method explicitly recovers known clinical morphologies that a neurophysiologist can directly validate. Its main advantage is that it works in the low-data regime where heavier models are a poor fit.

Abstract

Electroencephalography

Explore similar work

Interpretable EEG biomarkers with bag-of-waves: Spatial and temporal waveform dictionaries for low-data regimes

Comparing Post-Hoc Explainable AI Methods for Interpreting Black-Box EEG Models in Depression Detection

Hierarchic-EEG2Text: Assessing EEG-To-Text Decoding across Hierarchical Abstraction Levels