Article Preview
TopIntroduction
In the United States alone, there are approximately 250,000 medically refractory (i.e., interactable) epileptic patients (Elisevich et al., 1996). Previous studies on predictors of postoperative seizure freedom using multivariable analysis differ in both methodology and results, some including invasive tests or image modality data (Armon, Radtke, Friedman, & Dawson, 1996; Berg, Walczak, Hirsch, & Spencer, 1998; Clusmann et al., 2002; Hennessy et al., 2001; Janszky et al., 2005; Janszky et al., 2006; Jeha et al., 2006; Spencer et al., 2005; Tonini et al., 2004; Yun et al., 2006).
Uijl et al. (2008) proposed a model for prognosis after temporal lobe epilepsy surgery using a combination of predictors. The results of this study showed 85% seizure-freedom among patients with a high probability of seizure freedom, and 40% seizure-freedom among patients with a high risk of not becoming seizure-free. In another study, Antel et al. (2002) predicted surgical outcome in temporal lobe epilepsy patients using MRI and MRSI. In the Antel study, 75% of patients predicted accurately to fall in Engel class 1 after surgery on the basis of imaging alone. In the present study, we have developed a new algorithm based solely upon clinical, electrographic and neuropsychological data to predict surgical outcome. The main purpose of the study is to examine standard nonimaging evaluations as means of prognostication without relying on extraoperative electrocorticography (eECoG) or the many different imaging techniques used in these investigations. Clustering of patients in categorical groups and identification of the most effective features for each cluster permits the selection of a suitable classifier for each cluster. In order to identify the best features related to output (i.e., Engel classification), a combination of different feature selections and ranking algorithms are used aside a genetic algorithm search approach. For this purpose, integration of acquired multimodality data is necessary to obtain new effective features that can predict surgical outcome.
Data analysis with advanced computational methods (Siadat, Soltanian-Zadeh, Fotouhi, & Elisevich, 2005) leads to the use of artificial intelligence and machine learning algorithms. Accuracy of machine learning algorithms in an application depends on several factors. One of the challenges is selection of the best subset of available features. To this end, redundant features may be detected and eliminated. This may lead to more accurate, more effective, and faster results. Another challenge is selection of the best clustering approach. Our raw data include 139 instances with 102 attributes. These attributes are composed of video-scalp EEG analysis data based on well-defined standards, outcomes based on Engel classification, base tables (e.g., anti-epileptic drug information, seizure descriptions, medical history), neuropsychology, and Wada tests (Table 3). The large number of attributes may have negative effects on the classification accuracy. Therefore, effective and useful attributes should be found in order to make the classification simpler and more accurate.
In this paper, we introduce six different algorithms for feature selection, clustering and classification. The main differences among these algorithms reflect different feature selection and ranking approaches and clustering methods. The feature selection and attribute ranking algorithms are explained as are the different clustering approaches combined with a multilayer perceptron classifier to predict the outcome of epilepsy surgery based on Engel classification. The outputs of each feature selection algorithm are reported and the list of selected attributes for each cluster are shown in the supplementary tables.