Abstracts

Rationale: Language impairment is a common cognitive comorbidity in patients with temporal lobe epilepsy (TLE). Understanding the neurobiological substrates of language impairment may help to identify patients at greatest risk for post-operative language decline or who may be vulnerable to progressive language deficits. Here, we test the ability of an individualized marker of white matter (WM) network architecture, each patient’s structural connectome (SC), to discriminate between TLE patients who have impairment in naming and fluency from those with normal language profiles. Methods: Diffusion-weighted imaging, clinical, and neuropsychological data were obtained in 82 patients with TLE. Patients were classified as language impaired (TLE-LI; n=49) if they obtained scores >1.5 standard deviations below healthy controls on at least 2 of 3 neuropsychological measures of language function. Otherwise they were classified as non-language impaired (TLE-NLI; n = 33).Three models were created using XGBoost, a robust tree-based classifier, using patients’ a) clinical data, b) conventional atlas-derived WM tracts associated with language, and c) connection strengths derived from patients’ SCs. For the SC model, dimensionality reduction was achieved with principle component analysis. Model comparisons were validated using 1000 bootstrapped training samples. Results: The SC model outperformed both the clinical and tract models on prediction measures. In accuracy, the SC model (Accuracy: .80 +/- .02) significantly outperformed the tract model (Accuracy: .67 +/- .01) and the clinical model (Accuracy: .53 +/- .02; all p<.05). In other performance measures (see Table 1), the SC model also performed comparable to or better than the tract and clinical models, including superior performance on positive predictive value (PPV), specificity, and area under the curve (AUC).Next we characterized the specific connections within the SC network that contributed the most to model performance (see Figure 1). Connections were widely distributed and bilateral in nature (22 left-left connections; 19 left-right connections; 19 right-right connections). Overall, connections predominantly involved lateral temporal structures (40 connections to inferior, middle, or superior temporal gyri) rather than ventro-medial temporal connections (20 connections to parahippocampal, fusiform, or entorhinal cortex). Conclusions: We provide evidence that the pre-surgical SC can classify language impairment in TLE at the individual patient level. A range of inter- and intra-hemispheric connections predominantly centered on the lateral temporal lobe appear to underlie language impairments in naming and fluency, which cannot be inferred from clinical data or captured by conventional tract-based WM measures. These results emphasize the ability of the WM connectome to probe language network function in patients with TLE. Funding: CRM R01 NS065838