Standard frameworks to diagnose frontal lobe epilepsy prove nosologically unspecific, as they reveal deficits that also occur across other epilepsy types. Moreover, most existing results are undermined by low ecological validity. To face this challenge, we employed a naturalistic discourse paradigm combined with structural and functional brain connectivity measures in an analytical setting that integrates inferential statistics and machine learning. We assessed 19 frontal lobe epilepsy patients, 19 healthy controls, and 20 posterior cortex epilepsy patients, matched for sex, age, education, and neuropsychological variables. An ANCOVA revealed an interaction between group and condition for the action texts [F(2,110) = 8.14, P < .01, η2 = .26)], showing that patients with frontal lobe epilepsy were selectively impaired in grasping verb-related information. Such deficit was selectively and specifically correlated with (a) reduced integrity of the anterior thalamic radiation (r = 0.869, FDR-corrected P = .038), and (b) hypoconnectivity between the primary motor cortex and the left-parietal/supramarginal regions (r = 0.707, FDR-corrected P = .046). Machine-learning classifiers based on the above neurocognitive measures yielded 75% accuracy rates in discriminating individual frontal lobe epilepsy patients from both controls and posterior cortex epilepsy patients. Taken together, this multimodal approach opens new venues to complement mainstream cognitive assessments in epilepsy.