PyramidPat explainable feature engineering for multiclass electroencephalography psychiatric disorders: Explainable...

Area of Science:

• Neuroscience
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning models often neglect feature engineering, limiting the explainability of artificial intelligence (XAI) methods.
• Current research prioritizes accuracy over interpretability in machine learning applications.

Purpose of the Study:

• Introduce a novel explainable feature engineering (XFE) architecture, PyramidPat XFE, to enhance interpretability in machine learning.
• Develop an effective model for classifying psychiatric disorders using electroencephalogram (EEG) data.

Main Methods:

• The PyramidPat XFE pipeline involves four stages: PyramidPat feature extraction, INCA-based feature selection, tkNN classification, and DLob-based explanation.
• PyramidPat acts as a transformation-based feature extractor for multichannel signals.
• DLob converts selected feature identities into understandable lobe- and channel-based sentences for XAI.

Main Results:

• The proposed model achieved over 93% accuracy across seven test cases on a six-class EEG psychiatric disorder dataset.
• Leave-one-subject-out (LOSO) cross-validation demonstrated consistent high performance.
• The model successfully generated interpretable results, converting feature identities into descriptive sentences.

Conclusions:

• PyramidPat XFE is highly effective for EEG-based psychiatric disorder classification.
• The architecture provides compact and interpretable XAI outputs from complex EEG signals.
• This approach bridges the gap between high accuracy and explainability in machine learning for neurological applications.