Network phenotypes and their clinical significance in temporal lobe epilepsy using machine learning applications to morphological and functional graph theory metrics.

Machine learning analyses were performed on graph theory (GT) metrics extracted from brain functional and morphological data from temporal lobe epilepsy (TLE) patients in order to identify intrinsic network phenotypes and characterize their clinical significance. Participants were 97 TLE and 36 healthy controls from the Epilepsy Connectome Project. Each imaging modality (i.e., Resting-state functional Magnetic Resonance Imaging (RS-fMRI), and structural MRI) rendered 2 clusters: one comparable to controls and one deviating from controls. Participants were minimally overlapping across the identified clusters, suggesting that an abnormal functional GT phenotype did not necessarily mean an abnormal morphological GT phenotype for the same subject. Morphological clusters were associated with a significant difference in the estimated lifetime number of generalized tonic-clonic seizures and functional cluster membership was associated with age. Furthermore, controls exhibited significant correlations between functional GT metrics and cognition, while for TLE participants morphological GT metrics were linked to cognition, suggesting a dissociation between higher cognitive abilities and GT-derived network measures. Overall, these findings demonstrate the existence of clinically meaningful minimally overlapping phenotypes of morphological and functional GT networks. Functional network properties may underlie variance in cognition in healthy brains, but in the pathological state of epilepsy the cognitive limits might be primarily related to structural cerebral network properties.