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

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Epileptic seizure detection using Electroencephalography (EEG) signals is critical for patient care.
• Existing research often lacks robust uncertainty estimation and model calibration for EEG-based seizure classification.
• Accurate classification and reliable uncertainty quantification are essential for clinical decision-making.

Purpose of the Study:

• To investigate the combined use of Bayesian modeling, uncertainty estimation, and model calibration for improved EEG signal classification of epileptic seizures.
• To implement and evaluate representative Bayesian models including Gaussian Process, Bayesian Neural Network, and Monte-Carlo Dropout.
• To enhance the reliability and accuracy of epileptic seizure detection through refined probability outputs.

Main Methods:

• Implementation of Gaussian Process, Bayesian Neural Network, and Monte-Carlo Dropout for EEG signal analysis.
• Application of model calibration techniques to refine classification probabilities.
• Evaluation of the proposed methods on the Temple and Lemon EEG datasets.

Main Results:

• Demonstrated improvement in classification performance, evidenced by a 7.3% increase in Area Under the Curve (AUC).
• Achieved significant reductions in model error metrics: 38% decrease in negative log likelihood and 43% decrease in Brier score.
• Showcased enhanced predictive uncertainty estimation and model robustness for EEG data.

Conclusions:

• The proposed Bayesian approach effectively integrates uncertainty estimation and model calibration for accurate epileptic seizure classification from EEG signals.
• The methods provide more reliable probability estimates, crucial for clinical applications in epilepsy monitoring.
• This study highlights the potential of advanced Bayesian techniques to advance automated seizure detection systems.