DCPat-XFE: an explainable EEG model for psychogenic nonepileptic seizure detection

Area of Science:

• Neuroscience
• Medical Diagnostics
• Machine Learning in Healthcare

Background:

• Psychogenic Nonepileptic Seizures (PNES) are often misdiagnosed due to their similarity to epileptic seizures, necessitating improved diagnostic tools.
• Electroencephalography (EEG) is crucial for differentiating PNES from epilepsy, but current detection methods have limitations.
• The psychological origins of PNES require distinct diagnostic approaches compared to epilepsy's electrical origins.

Purpose of the Study:

• To introduce a novel explainable feature-engineering (XFE) model for detecting PNES using EEG.
• To curate a specialized PNES EEG dataset with expert annotations for Normal, PNES with Verbal Suggestion Provocation (VSP+), and PNES without VSP (VSP-).
• To evaluate the performance and interpretability of the proposed XFE framework in classifying different EEG patterns.

Main Methods:

• Developed an XFE framework comprising Distance Counter Pattern (DCPat) for feature extraction, Cumulative Weight-based Neighborhood Component Analysis (CWNCA) for selection, t-algorithm k-Nearest Neighbors (tkNN) with Iterative Majority Voting (IMV) for classification, and Directed Lobish (DLob) for interpretation.
• Utilized a curated EEG dataset with expert-labeled Normal, PNES VSP+, and PNES VSP- classes across four distinct case studies.
• Employed DLob for symbolic interpretation and cortical connectome mapping to explain PNES-related EEG patterns.

Main Results:

• The DCPat XFE framework achieved over 96.5% accuracy across all four evaluated cases.
• Case 2 (Normal vs. PNES VSP-) demonstrated the highest accuracy at 99.11%.
• DLob outputs provided clear symbolic explanations and connectome diagrams, enhancing the model's interpretability.

Conclusions:

• The novel DCPat-based XFE framework offers a highly accurate and interpretable method for PNES detection from EEG data.
• The model's ability to provide clear symbolic explanations supports its potential as a reliable clinical decision support tool.
• This research contributes a valuable dataset and a robust methodology for advancing PNES diagnosis.