Uncertainty Estimation and Model Calibration in EEG Signal Classification for Epileptic Seizures Detection

This paper studies Bayesian modeling, uncertainty estimation and model calibration for Electroencephalography (EEG) signal classification. Prior research lacks studies that combine uncertainty estimation with model calibration in EEG data analysis for epileptic seizures. In this work, we implement the Gaussian process, the Monte-Carlo dropout and the Bayesian neural network as three representative Bayesian models. The Gaussian Process offers a flexible non-parametric framework for capturing underlying patterns in EEG data, while the Bayesian Neural Network and MC Dropout enhance predictive uncertainty estimation and model robustness. Moreover, the model calibration technique is employed to refine the final probabilities, ensuring improved reliability of the classification outcomes. Through evaluation on Temple and Lemon EEG datasets, the proposed approach shows promising results, i.e., 7.3% improvement in area under curve, 38% reduction in negative log likelihood and 43% reduction in the Brier score, demonstrating its potential in accurate uncertainty estimation and calibrated EEG epileptic signal classification.