Operational Transformer: An investigation of epilepsy detection

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Electroencephalography (EEG) signals record brain electrical activity, vital for diagnosing conditions like epilepsy.
• Interpreting complex EEG data remains a significant challenge, necessitating advanced analytical methods.

Purpose of the Study:

• To introduce a novel transformer model, the Operational Transformer (OpT), for multichannel EEG signal classification.
• To present a new explainable feature engineering (XFE) framework to evaluate the OpT model's classification capabilities.

Main Methods:

• Feature derivation using OpT and a transition table feature extractor.
• Identification of significant features via cumulative weighted iterative neighborhood component analysis (CWINCA).
• Classification using k-nearest neighbors (kNN) and explainable output generation with the Directed Lobish (DLob) method.

Main Results:

• The OpT-driven XFE model achieved 99.99% accuracy with 10-fold cross-validation and 84.74% with leave-one-subject-out cross-validation on an epilepsy dataset.
• A connectome diagram was successfully generated using DLob for interpretability.
• High classification performance and interpretable results were demonstrated.

Conclusions:

• The OpT-driven XFE model offers a powerful approach for EEG signal classification, particularly for epilepsy diagnosis.
• The framework provides significant contributions to both feature engineering and neuroscience through high performance and interpretability.