[Paper Review] Random Forest classifier for EEG-based seizure prediction

Remy Ben Messaoud, Mario Chávez|arXiv (Cornell University)|Jun 2, 2021

EEG and Brain-Computer Interfaces32 references4 citations

TL;DR

This study proposes a Random Forest classifier for EEG-based seizure prediction, leveraging epoch-wise probability estimation of pre-ictal states to improve accuracy and reduce false positives. On the CHB-MIT scalp EEG dataset with a 5-minute seizure prediction horizon, it achieves 82.07% sensitivity and a low false positive rate of 0.0799/h, outperforming state-of-the-art methods.

ABSTRACT

Epileptic seizure prediction has gained considerable interest in the computational Epilepsy research community. This paper presents a Machine Learning based method for epileptic seizure prediction which outperforms state-of-the art methods. We compute a probability for a given epoch, of being pre-ictal against interictal using the Random Forest classifier and introduce new concepts to enhance the robustness of the algorithm to false alarms. We assessed our method on 20 patients of the benchmark scalp EEG CHB-MIT dataset for a seizure prediction horizon (SPH) of 5 minutes and a seizure occurrence period (SOP) of 30 minutes. Our approach achieves a sensitivity of 82.07 % and a low false positive rate (FPR) of 0.0799 /h. We also tested our approach on intracranial EEG recordings.

Motivation & Objective

• To develop a machine learning-based method for reliable epileptic seizure prediction using EEG signals.
• To improve upon existing state-of-the-art approaches by reducing false positive rates while maintaining high sensitivity.
• To enhance algorithmic robustness against false alarms through novel feature engineering and classifier design.
• To validate the method on both scalp and intracranial EEG recordings for broader clinical applicability.

Proposed method

• The Random Forest classifier is trained to predict the probability of a given EEG epoch being pre-ictal versus interictal.
• New features are engineered to capture temporal and spectral dynamics in EEG signals, improving discriminative power.
• The model uses a sliding window approach to process continuous EEG data into discrete epochs for classification.
• A probability thresholding strategy is applied to balance sensitivity and false positive rate.
• The method incorporates ensemble learning to enhance generalization and robustness across diverse patient data.
• The approach is validated on the benchmark CHB-MIT scalp EEG dataset and tested on intracranial EEG recordings for cross-modality performance.

Experimental results

Research questions

• RQ1 Can a Random Forest classifier achieve higher sensitivity and lower false positive rates than existing methods in EEG-based seizure prediction?
• RQ2 How does the inclusion of novel EEG features impact the robustness of seizure prediction models?
• RQ3 To what extent does the model generalize across different EEG modalities, such as scalp and intracranial recordings?
• RQ4 How effective is the probability-based classification framework in minimizing false alarms while maintaining high sensitivity?
• RQ5 What is the performance of the model under a 5-minute seizure prediction horizon and 30-minute seizure occurrence period?

Key findings

• The proposed method achieves a sensitivity of 82.07% on the CHB-MIT scalp EEG dataset with a 5-minute seizure prediction horizon.
• The false positive rate is reduced to 0.0799 per hour, representing a significant improvement over prior state-of-the-art methods.
• The model demonstrates strong generalization capability, achieving comparable performance on intracranial EEG recordings.
• The integration of new feature engineering techniques enhances the classifier's ability to detect subtle pre-ictal patterns in EEG signals.
• The probability-based output of the Random Forest enables tunable trade-offs between sensitivity and false positive rate.
• The method outperforms existing approaches in both sensitivity and false positive rate metrics under the same evaluation protocol.