A Comprehensive Machine-Learning-Based Software Pipeline to Classify EEG Signals: A Case Study on PNES vs. Control...

Area of Science:

• Neurology
• Biomedical Engineering
• Machine Learning

Background:

• Diagnosing psychogenic nonepileptic seizures (PNES) is challenging for neurologists due to the lack of clear electrophysiological biomarkers.
• Current diagnostic methods rely on video electroencephalography (EEG) monitoring and clinical history, which can be time-consuming and complex.
• There is a need for objective, data-driven tools to support the clinical diagnosis of PNES.

Purpose of the Study:

• To develop and evaluate a machine learning (ML) pipeline for classifying EEG segments from PNES patients and healthy controls (CNT).
• To assess the effectiveness of statistical features extracted from power spectral density (PSD) maps in discriminating PNES.
• To compare the performance of different supervised ML algorithms (SVM, LDA, BN) for PNES detection using EEG time series.

Main Methods:

• A semiautomatic signal processing technique was employed to extract statistical features (mean, standard deviation, kurtosis, skewness) from EEG power spectral density (PSD) maps across five frequency bands (delta, theta, alpha, beta, and 1-32 Hz).
• Feature vectors were used to train and test three supervised ML classifiers: Support Vector Machine (SVM), Linear Discriminant Analysis (LDA), and Bayesian Network (BN).
• The ML pipeline's performance was validated using Random Split (RS) and Leave-One-Out (LOO) cross-validation on a dataset of 20 EEG recordings (10 PNES, 10 CNT).

Main Results:

• The ML pipeline achieved high classification accuracy, with RS-SVM reaching 0.97 ± 0.013 and RS-LDA achieving 0.99 ± 0.02.
• Leave-One-Out validation demonstrated comparable performance, with LOO-SVM at 0.98 ± 0.0233 and LOO-LDA at 0.98 ± 0.124.
• Bayesian Network (BN) showed lower accuracy (RS-BN: 0.82 ± 0.109, LOO-BN: 0.81 ± 0.109) compared to SVM and LDA.

Conclusions:

• The proposed data-driven ML pipeline, particularly using SVM and LDA, demonstrates significant potential in accurately discriminating between PNES and healthy controls based on EEG data.
• The findings suggest that this ML approach can serve as a valuable supplementary tool for clinicians in the diagnosis of PNES.
• Further research and validation on larger datasets are warranted to fully integrate this technology into routine clinical practice.