Anterior temporal lobe resection (ATLR), while successful can result in lasting impairment of language function. White matter bundles have been shown to explain some of the variance seen in language decline after ATLR. Network analysis of the structural connectome has been shown superior in predicting preoperative language ability but remains unexplored in predicting postoperative ability.

Diffusion MRI-based tractography was used to generate the preoperative connectome on 54 left-lat- eralised (as determined by functional MRI), left-hemisphere ATLR. Postoperative connectomes were estimated via manually drawn resection masks. Graded naming test (GNT), semantic, and letter fluency were binarised into significant decline or not (via their reliable change indices). Strength (sum of connec- tions) and betweenness centrality (interconnectivity) network changes were generated using pre- and postoperative connectomes as predictor variables. Each model was entered into a linear support vector machine incorporating synthetic minority over-sampling technique for class imbalances.

Strength changes alone accurately predicted 81.6% of patients who had GNT decline. Betweenness centrality changes accurately predicted 73.3% of patients who had letter fluency decline. Patients with semantic decline were predicted equally as well by strength and betweenness centrality changes (accuracy=71.1%).

These findings demonstrate the usefulness of the structural network in predicting and potentially prevent- ing postoperative language decline.