Machine learning–XGBoost analysis of language networks to classify patients with epilepsy

Our goal was to apply a statistical approach to allow the identification of atypical language patterns and to differentiate patients with epilepsy from healthy subjects, based on their cerebral activity, as assessed by functional MRI (fMRI). Patients with focal epilepsy show reorganization or plasticity of brain networks involved in cognitive functions, inducing 'atypical' (compared to 'typical' in healthy people) brain profiles. Moreover, some of these patients suffer from drug-resistant epilepsy, and they undergo surgery to stop seizures. The neurosurgeon should only remove the zone generating seizures and must preserve cognitive functions to avoid deficits. To preserve functions, one should know how they are represented in the patient's brain, which is in general different from that of healthy subjects. For this purpose, in the pre-surgical stage, robust and efficient methods are required to identify atypical from typical representations. Given the frequent location of regions generating seizures in the vicinity of language networks, one important function to be considered is language. The risk of language impairment after surgery is determined pre-surgically by mapping language networks. In clinical settings, cognitive mapping is classically performed with fMRI. The fMRI analyses allowing the identification of atypical patterns of language networks in patients are not sufficiently robust and require additional statistic approaches. In this study, we report the use of a statistical nonlinear machine learning classification, the Extreme Gradient Boosting (XGBoost) algorithm, to identify atypical patterns and classify 55 participants as healthy subjects or patients with epilepsy. XGBoost analyses were based on neurophysiological features in five language regions (three frontal and two temporal) in both hemispheres and activated with fMRI for a phonological (PHONO) and a semantic (SEM) language task. These features were combined into 135 cognitively plausible subsets and further submitted to selection and binary classification. Classification performance was scored with the Area Under the receiver operating characteristic curve (AUC). Our results showed that the subset SEMLH BA₄7-21 (left fronto-temporal activation induced by the SEM task) provided the best discrimination between the two groups (AUC of 91 ± 5%). The results are discussed in the framework of the current debates of language reorganization in focal epilepsy.