KnowEEG: Explainable Knowledge Driven EEG Classification

The gist: KnowEEG introduces a novel explainable machine learning approach for EEG classification that achieves performance comparable to or exceeding state-of-the-art deep learning models across five different classification tasks while providing inherent explainability through feature importance scores.

Problem and Approach

Electroencephalography (EEG) is a method of recording brain activity that shows significant promise in applications ranging from disease classification to emotion detection and brain-computer interfaces EEG classification is a challenging task given the low signal-to-noise ratio in EEG signals and the presence of inter-subject variability requiring models to generalize well across individuals Traditional machine learning methods for EEG classification have relied on standard hand-crafted features as inputs to models such as support vector machines, random forests or neural networks Deep learning models are often opaque and their inner workings are hard to interpret KnowEEG extracts a comprehensive set of per-electrode features, filters them using statistical tests, and integrates between-electrode connectivity statistics These features are then input to our modified Random Forest model (Fusion Forest) that balances per electrode statistics with between electrode connectivity features in growing the trees of the forest By incorporating knowledge from both the generalized time-series and EEGspecific domains, KnowEEG achieves performance comparable to or exceeding state-of-the-art deep learning models across five different classification tasks: emotion detection, mental workload classification, eyes open/closed detection, abnormal EEG classification, and event detection In addition to high performance, KnowEEG provides inherent explainability through feature importance scores for understandable features

KnowEEG Model Architecture

The KnowEEG pipeline is structured into two parallel threads to capture different aspects of the EEG data Thread One involves calculating per electrode features Per electrode we calculate 783 features from both the generalized time series domain and the EEG-specific domain These per electrode features are concatenated after being calculated For binary classification we use the Mann-Whitney U test while for multi-class classification Kruskal–Wallis tests are applied Thread Two calculates between electrode connectivity features We select Spearman and Pearson correlation, which measure linear and non7 linear correlation between electrodes For our pipeline we select Coherence (Coh), Imaginary Coherence (ImCoh), Pairwise Phase Consistency (PPC), Phase Lag Value (PLV), Phase Lag Index (PLI), Directed Phase Lag Index (DPLI) and Weighted Phase Lag Index (WPLI) [25,page8> Users of the pipeline can select any number of connectivity metrics that they expect could be informative for classification Feature Fusion combines per-electrode and between electrode features using the Fusion Forest Algorithm 1 The Fusion Forest Algorithm balances two feature modes (Xa and Xb) by independently selecting a random subset of features from each (lines 5-6) before fitting each of the K trees We propose a modification of the traditional Random Forest algorithm so that for each tree a random subset of features is selected from mode 1 and separately from mode 2

Experimental Results and Performance

KnowEEG was tested across five publicly available EEG datasets Table 2 shows the average performance for all models across the DREAMER, Crowdsourced, STEW, TUEV and TUAB datasets With 10 total performance metrics across 5 datasets KnowEEG achieves best performance vs. all other models on 5/10 metrics KnowEEG is the best performing model overall versus state-of-the-art competitors On Crowdsourced, KnowEEG has the second highest accuracy with 92.81 % performing marginally worse than the pre-trained EEG2Rep model On STEW with the task of mental workload classification KnowEEG significantly outperforms all other models in both accuracy and AUROC with standard deviations of under 1 for both metrics again illustrating consistent performance across seeds For the TUEV dataset KnowEEG is third in balanced accuracy yet significantly outperforms all other models in weighted F1 score EEG2Rep Pre-trained [9] is overall the second best performing model achieving best performance or second best performance on 8/10 metrics across the five datasets KnowEEG has the significant advantage of not requiring GPU for training

Explainability Analysis

Section 4.5 illustrates the explainability properties of KnowEEG using the Crowdsource dataset and demonstrates key explainability benefits In hyperparameter selection, Functional Power Connectivity (FPC) was selected as the connectivity measure for the Crowdsource dataset We have uncovered known neuroscience knowledge. Specifically, we have found increased mean alpha band activity and connectivity in the eyes closed state compared to eyes open as well as increased mean gamma band activity in the eyes open class This ability of our pipeline to uncover knowledge about the classes is a significant strength We find that Occipital Lobe signal statistics aswell as gamma band activity, alpha band activity and alpha band connectivity to be discriminative of the eyes open and eyes closed classes From Table 3 it is notable that 7 out of the top 7 highest importance features are from the Occipital brain region This suggests that the Occipital brain region provides features that are most discriminative of the two classes We have verified this versus existing neuroscience literature

Conclusion and Future Work

We present KnowEEG, an explainable high-performance pipeline for EEG classification We demonstrate through extensive experimentation across five different EEG classification tasks that KnowEEG outperforms state-of-the-art deep learning methods KnowEEG has the added benefits of not requiring GPUs for training and is also explainable Future work will explore additional EEG datasets to determine if the performance of KnowEEG remains high We should deploy this pipeline on real world data in the healthcare domain to determine if the pipeline can successfully discriminate between classes where the differences between the classes are not already understood This could be particularly impactful in domains such as neurodegenerative disease classification where EEG is already being explored as a means of classification Further analysis in future work should be done on why KnowEEG outperforms competitors We could determine what the extent of the shared information between learnt self-supervised representations such as in EEG2Rep [9] and the features in KnowEEG

Funding

This work was supported by UK Research and Innovation grant EP/S022937/1: Interactive Artificial Intelligence The references section lists numerous works supporting the methodology and results of the paper The appendix provides technical details for the setup of KnowEEG including package versions used and specific settings for feature calculation Appendix D details the preprocessing steps applied to each dataset, such as bandpass filtering and windowing The paper concludes by summarizing its findings regarding performance superiority and explainability across various EEG classification domains The work highlights the advantage of using CPU-based feature calculation over GPU-intensive deep learning models for accessibility The overall contribution lies in creating a unified, high-performance, explainable pipeline applicable across multiple EEG classification domains This approach successfully combines generalized time series knowledge with EEG-specific connectivity measures to achieve superior results compared to existing methods The paper confirms that the discovered knowledge about the classes is consistent with established neuroscience literature regarding alpha and gamma band activity differences between eyes closed and eyes open states This capability allows KnowEEG to aid in the discovery of new information where class differences are not yet understood The pipeline's performance across five different datasets confirms its robustness across diverse EEG classification tasks The paper successfully bridges the gap between high classification performance and inherent model interpretability in the field of EEG analysis

--- Page 1 ---

KnowEEG: Explainable Knowledge Driven EEG Classification Amarpal Sahotaa, Navid Mohammadi Foumanib, Raul Santos-Rodrigueza, Zahraa S.

1. Bold Feature Filtering for Robustness: Improve feature relevance by filtering features using univariate nonparametric hypothesis tests (e.g., Mann-Whitney U test for binary targets and Kruskal-Wallis tests for multiclass targets), with adjusted p-values using the Benjamini–Yekutieli correction to control the false discovery rate (α = 0.05). This prevents the model from being overwhelmed by uninformative features, leading to a more focused and robust classifier.

2. Bold Feature Fusion Strategy: Implement Fusion Forest where each tree in the Random Forest selects a random subset of features from mode 1 and separately from mode 2, as proposed by modifying the traditional Random Forest algorithm. This prevents the feature set with a higher dimension from dominating the model, ensuring that both generalized time-series statistics and between-electrode connectivity features contribute equally to every decision tree.

3. Bold Connectivity Metric Selection: Systematically allow users to select any number of connectivity metrics as a hyperparameter for Thread Two, instead of selecting a single metric at a time. This enables the system to test combinations of measures (like Coherence, PLI, and FPC) to find the optimal configuration for specific classification tasks.

4. Bold Task-Specific Feature Importance Discovery: Leverage the explainability analysis in Section 4.5 to discover domain-specific knowledge; specifically, we have uncovered known neuroscience knowledge, such as increased mean alpha band activity and connectivity in the eyes closed state compared to eyes open. This allows the system to verify or suggest hypotheses aligned with existing neurophysiological literature.

5. Bold Computational Efficiency for Deployment: Utilize the fact that KnowEEG does not require any GPU resource for training and is primarily CPU-based, as noted in Section 4.6. This makes the pipeline more accessible for deployment on resource-constrained edge devices or real-world clinical settings where GPU access is limited.