Algorithmic Analysis of Clinical and Neuropsychological Data in Localization-Related Epilepsy

Abstract

The current study examines algorithmic approaches for analysis of nonimaging (i.e., clinical, electrographic and neuropsychological) attributes in localization-related epilepsy (LRE), specifically, their impact on the selection of patients for surgical consideration. Both invasive electrographic and imaging data are excluded here to concentrate upon the initial clinical presentation and the varied elements of the seizure history, ictal semiology, risk and seizure-precipitating factors and physical findings in addition to several features of the neuropsychological profile including various parameters of cognition and both speech and memory lateralization. The data was accrued in a database of temporal lobe epilepsy patients (HBIDS). Six algorithms comprising feature selection, clustering and classification approaches were used. The Correlation-Based Feature Selection (CFS) and the Classifier Subset Evaluator (CSE) with the Genetic Algorithm (GA) search tool and ReliefF Attribute Evaluation approaches provided for feature selection. The Expectation Maximization (EM) Class Clustering and Incremental Conceptual Clustering (COBWEB) provided data clustering and the Multilayer Perceptron (MLP) Classifier was the classification tool at all stages of the study. The Engel Classification was used as an output of classifier for surgical success. Attributes demonstrating the highest correlation with the outcome class and the least intercorrelation with each other, according to CFS, were selected. These were then ranked using ReliefF and the top rankings chosen. The best attribute combination for each cluster was found by MLP. COBWEB provided the best results showing an association of 56% with Engel class. In conclusion, an algorithmic approach to the study of LRE is feasible with current findings supporting the need for correlative electrographic and imaging data and a greater archival population.