Benchmarking Machine Learning Models for Seizure Classification on the BEED Dataset
A comprehensive benchmarking study comparing six machine learning architectures (CatBoost, XGBoost, LightGBM, Random Forest, SVM, Logistic Regression) for EEG seizure classification with noise-injection robustness stress tests and SHAP biomarker interpretability.
THE QUESTION
Which machine learning architectures achieve the optimal trade-off between predictive classification accuracy and empirical robustness against Gaussian noise and data perturbations in clinical EEG seizure detection?
BACKGROUND
Epilepsy affects over 50 million individuals globally, requiring rapid and accurate electroencephalography (EEG) seizure detection. While numerous machine learning models achieve high accuracy on curated benchmarks, their clinical viability hinges on robustness—their ability to maintain performance under sensor noise, motion artifacts, and class imbalance. This research benchmarked six models on the Bonn EEG Epilepsy Dataset (BEED), implementing stress tests and SHAP feature attribution to determine clinical deployment reliability.
APPROACH OVERVIEW
EEG SIGNAL EXTRACTION & PREPROCESSING
Ingestion of 500 EEG segments (Sets A–E) from the Bonn dataset, applying bandpass filtering (0.5–100 Hz) and notch filtering (50 Hz powerline interference).
MULTI-DOMAIN FEATURE ENGINEERING
Extraction of 45 physiological features per channel: time-domain statistics, frequency-domain spectral power (delta, theta, alpha, beta, gamma), and nonlinear dynamics (Hjorth parameters, Hurst exponent, approximate entropy).
MODEL BENCHMARKING & TUNING
Systematic 5-fold stratified cross-validation across Logistic Regression, SVM-RBF, Random Forest, XGBoost, LightGBM, and CatBoost with Bayesian hyperparameter optimization.
ROBUSTNESS STRESS TESTING
Quantitative degradation profiling under synthetic Gaussian noise injection (sigma = 5%, 10%, 15%), sample dropout, and simulated minority-class imbalance.
EXPLAINABLE AI & SHAP ATTRIBUTION
Game-theoretic SHAP feature attribution validating whether top predictive weights correspond to established neurological biomarkers (alpha/beta power ratio, spectral entropy).
METHODS
- •Dataset: Bonn EEG Epilepsy Dataset (BEED) comprising 500 single-channel EEG segments across five clinical conditions (surface healthy eyes open/closed, intracranial hippocampal/epileptogenic zone, and ictal seizure events).
- •Feature Extraction: 45 multi-domain features encompassing time-domain metrics (zero-crossing rate, line length, kurtosis), band power spectral densities, and nonlinear chaos indicators (approximate entropy, Hurst exponent, Hjorth mobility/complexity).
- •Cross-Validation: Stratified 5-fold cross-validation with variance tracking to ensure generalizability across partitioned clinical sets.
- •Robustness Stress Testing: Controlled perturbation experiments evaluating accuracy degradation under 5%–15% Gaussian noise injection, 10%–20% random sample dropout, and feature corruption.
- •Model Interpretability: TreeSHAP computation to extract global and local feature importance rankings, correlating machine learning weights with clinical EEG biomarkers.
KEY DATA SNAPSHOT
RESULTS
CatBoost achieved the highest classification performance (97.5% accuracy, 0.974 F1-score, 0.996 ROC-AUC) with minimal cross-validation variance (+/-0.6%). In stress testing, CatBoost demonstrated superior stability, suffering only a 1.5% drop under 10% Gaussian noise compared to 3.8% in SVM and 4.2% in Logistic Regression. SHAP analysis confirmed alpha/beta power ratio and spectral entropy as the primary drivers of seizure identification.
DISCUSSION
Gradient-boosted decision trees (CatBoost, XGBoost) consistently outperformed kernel and linear baselines, reflecting their capacity to capture nonlinear interactions among multiscale frequency bands.
CatBoost demonstrated remarkable resilience to synthetic noise perturbations (-1.5% degradation), highlighting symmetric tree structures as a protective regularizer against sensor artifacts.
SHAP feature attribution successfully aligned computational feature importance with clinical neurophysiology: alpha/beta suppression and high spectral entropy were identified as the strongest markers of ictal transitions.
LIMITATIONS
- •The BEED dataset utilizes pre-segmented, artifact-filtered single-channel recordings; prospective real-time continuous multi-channel scalp EEG signals present higher noise dimensionality.
- •Binary classification (ictal vs. non-ictal) does not resolve specific seizure semiology or distinct epileptic syndrome sub-types.
IMPACT & APPLICATION
DATA & REPRODUCIBILITY
Analytical code and specific target coordinates are currently held under institutional review and confidential protocol.
REFERENCES
- andrzejak2001Andrzejak, R. G. et al. Indicating nonlinear deterministic and finite-dimensional structures in time series of normal and epileptic brain activity. Phys. Rev. E 64, 061907 (2001).
- lundberg2017Lundberg, S. M. & Lee, S. I. A unified approach to interpreting model predictions. Adv. Neural Inf. Process. Syst. 30, 4765–4774 (2017).