An Explainable Feature Engineering Model Based on Automata Pattern: Investigations on the EEG Artifact...

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Electroencephalography (EEG) analysis requires robust feature extraction methods.
• Existing methods often lack interpretability, hindering clinical application.
• Explainable AI (XAI) is crucial for understanding complex biological data.

Purpose of the Study:

• Introduce Automata Pattern (AutPat) as a novel EEG feature extractor.
• Develop an explainable feature engineering (XFE) pipeline for EEG analysis.
• Evaluate AutPat's performance and interpretability across diverse EEG tasks.

Main Methods:

• AutPat feature extraction from raw EEG data.
• Feature selection using cumulative weighted iterative neighborhood component analysis (CWINCA).
• Classification via a t-algorithm-based k-nearest neighbors (tkNN) classifier.
• Interpretability through Directed Lobish (DLob) symbols and connectome diagrams.

Main Results:

• AutPat-based XFE achieved >88% classification accuracy on artifact, stress, and mental performance detection tasks.
• CWINCA effectively reduced feature dimensionality while preserving classification accuracy.
• The DLob layer generated symbolic outputs and 8x8 cortical connectome matrices for interpretability.

Conclusions:

• AutPat, integrated into an XFE pipeline with CWINCA and tkNN, provides a compact, accurate, and interpretable EEG analysis solution.
• The pipeline demonstrates practical utility for EEG analysis requiring both high performance and inherent symbolic explanations.
• AutPat-based XFE represents a significant advancement for explainable EEG data interpretation.