Seizure Type Classification Based on Hybrid Feature Engineering and Mutual Information Analysis Using...

Area of Science:

• Neuroscience
• Medical Informatics
• Machine Learning

Background:

• Epilepsy diagnosis is challenged by diverse seizure types and subjective manual interpretation of electroencephalogram (EEG) data.
• Automated, accurate seizure classification is crucial for improving patient outcomes, especially with imbalanced datasets.

Purpose of the Study:

• To develop a hybrid framework for automated multi-class seizure type classification using EEG signals.
• To enhance classification precision and address data challenges through segment-wise processing and multi-band feature engineering.

Main Methods:

• EEG signals from the TUSZ dataset were segmented, and multi-band features (statistical, entropy, wavelet, Hurst, Hjorth) were extracted.
• Mutual Information (MI) was used for optimal feature selection, and seven machine learning models were evaluated using 10-fold cross-validation with class balancing.
• The XGBoost model was identified as the top performer.

Main Results:

• XGBoost achieved the highest performance with an accuracy of 0.8710, F1 score of 0.8721, and AUC of 0.9797.
• Gamma-band features were found to be the most important.
• Confusion matrices revealed robust discrimination, though some overlap was observed in focal seizure subtypes.

Conclusions:

• The developed hybrid framework effectively integrates multi-band features and MI for advanced seizure type classification.
• This approach provides a scalable and interpretable tool to support clinical epilepsy diagnostics.