Brain Connectivity Analysis on Long-term Electroencephalograms in Pediatric Epileptic Patients with Intractable Seizures

Epilepsy is one of the most common neurological disorders in children and is characterized by recurrent seizures resulting from abnormal neuronal activity. In particular, in patients with drug-resistant epilepsy, prolonged electroencephalographic (EEG) monitoring provides an important source of information for investigating the electrophysiological dynamics associated with seizure generation. In this context, quantitative EEG analysis, particularly the assessment of functional connectivity and brain network organization, may contribute to the identification of alterations occurring before seizure onset and to a better characterization of the preictal state. The aim of this thesis was to investigate changes in functional brain connectivity and network organization preceding seizure onset in pediatric patients with drug-resistant epilepsy. Long-term scalp EEG recordings from the CHB-MIT database were analyzed. After preprocessing, EEG signals were divided into non-overlapping 4-s windows and classified according to their temporal relationship with seizures. The preictal state was defined as the 30 min preceding seizure onset, whereas interictal activity was required to occur at least 4 h from any seizure. Artifact-contaminated windows were identified using a Joint Probability approach and excluded from subsequent analyses. Functional connectivity was characterized using 32 configurations, including broadband and frequency-specific measures based on linear correlation, information theory, phase synchronization, coherence, and amplitude coupling. A subset of 21 non-negative and undirected connectivity configurations was subsequently used to construct weighted functional networks and extract global and nodal graph-theoretical metrics. Preictal and interictal conditions were compared at the subject level using paired statistical analyses with correction for multiple comparisons. Finally, functional-connectivity and global graph-theoretical features were evaluated for preictal–interictal classification using both preliminary five-fold cross-validation and subject-independent leave-one-subject-out (LOSO) validation. The functional-connectivity analysis did not identify any connections that remained statistically significant after False Discovery Rate correction, and no consistent temporal trend was observed across the 30-min preictal interval. Similarly, none of the global graph-theoretical metrics showed significant preictal–interictal differences after correction. In contrast, three nodal betweenness-centrality comparisons remained significant, indicating localized and frequency-dependent alterations in network organization. In the preliminary five-fold cross-validation, the best-performing functional-connectivity classifier was a Fine Gaussian SVM, achieving an accuracy of 84.0%, whereas a Fine KNN achieved 77.5% accuracy using global graph-theoretical features. However, performance decreased under subject-independent LOSO validation. Functional-connectivity features achieved a balanced accuracy of 54.22% and a ROC-AUC of 0.533, whereas global graph-theoretical features achieved a balanced accuracy of 51.20% and a ROC-AUC of 0.512. Taken together, these findings indicate that the preictal state in pediatric drug-resistant epilepsy was not associated with a homogeneous alteration in pairwise functional connectivity or global network organization across subjects. Instead, selected localized changes in network topology were identified, suggesting that preictal network dynamics may be spatially and frequency dependent. The limited subject-independent classification performance further highlights the marked inter-subject variability of pediatric epilepsy and supports the investigation of more individualized approaches for the characterization of preictal EEG dynamics.