Genetic Programming and Frequent Itemset Mining to Identify Feature Selection Patterns of iEEG and fMRI Epilepsy Data

Area of Science:

• Neuroscience and Biomedical Engineering
• Epilepsy Research and Diagnostics
• Signal Processing and Machine Learning

Background:

• Pattern classification of intracranial electroencephalogram (iEEG) and functional magnetic resonance imaging (fMRI) signals is crucial for understanding epilepsy origins and localizing dysfunctional brain tissue.
• Genetic programming (GP) has shown effectiveness in implicitly selecting features for discerning interictal activity in iEEG and fMRI data, outperforming conventional methods.
• Uncertainty remains regarding the consistency and generalizability of GP-derived features across different patient datasets for both iEEG and fMRI modalities.

Purpose of the Study:

• To investigate the reproducibility of implicitly selecting features using a genetic programming (GP) algorithm for classifying interictal activity in epilepsy patients.
• To assess the within-subject consistency and across-subject variability of GP-selected features for both iEEG and fMRI data.
• To evaluate the potential need for patient-specific feature selection and classification strategies in epilepsy diagnostics.

Main Methods:

• Multiple feature selection trials were conducted using a genetic programming (GP) algorithm on separate iEEG and fMRI epilepsy patient datasets.
• Frequent itemset mining (FIM) was employed following GP-based feature selection to analyze the identified patterns.
• Nearest-neighbor classification was applied, utilizing data from 30 GP generations for performance evaluation.

Main Results:

• Observed within-subject consistency in feature selection, indicating reliable results within an individual's data.
• Demonstrated across-subject variability in selected features, suggesting a need for personalized approaches in epilepsy analysis.
• Achieved over 60% median sensitivity and selectivity for fMRI, and over 65% median sensitivity and selectivity for iEEG (with one exception), showcasing the efficacy of GP-based classification.

Conclusions:

• Implicit feature selection with GP exhibits within-subject consistency but across-subject variability for iEEG and fMRI epilepsy data.
• The findings highlight a clear need for patient-specific features and potentially patient-specific feature selection or classification methods.
• GP-based analysis, combined with nearest-neighbor classification, provides promising diagnostic performance for both iEEG and fMRI modalities in epilepsy.