Postdoctoral Fellow University of Pennsylvania, Pennsylvania
Rationale: Language impairment is a prominent complaint of patients with left temporal lobe epilepsy (TLE). Disruptions of static and dynamic functional network organization, altered fronto-temporal task-related functional MRI activation and impaired white matter fasciculi underlie these neuropsychological symptoms1, and point to distributed abnormalities of neural functioning and energetic metabolism. Indeed, ipsilateral temporal hypometabolism is commonly observed in TLE patients, while its relevance to language dysfunction is overlooked. Here, we apply a feature selection method based on Random Forest (RF)2 to regional laterality indices (LI) of brain metabolism measured with FDG-PET in left TLE patients, to identify regional metabolic asymmetries that may inform our understanding of the neural substrates of language dysfunction. Methods: Thirty-five left TLE patients underwent both structural MRI (T1) and interictal, resting-state FDG-PET scans as part of presurgical evaluation. The PET image was coregistered to the T1, spatially smoothed and intensity normalized by the global mean. Using a symmetric version of the Lausanne atlas inversely warped onto each subject’s T1 space, we estimated regional metabolic values from 122 regions of interest (ROIs; 61 pairs). Regional gray matter volumes (GMV) from the same ROIs were extracted via the Computational Anatomy Toolbox (CAT12). The LI [(L-R)/(L+R)] of both the metabolism and the GMV was computed for all ROI pairs as predictors, as well as demographic and clinical predictors including age, sex, handedness, education, age at epilepsy onset, duration, MRI evidenced pathology, history of FBTCS, focal seizure frequency, AED load, and total brain volume. Verbal fluency, as a linguistic measure, was assessed via the Controlled Oral Word Association (COWA) and Animal Naming (AN) tests. Separately using COWA and AN as response variables, we built RF models to evaluate the variable importance of all the aforementioned predictors and select the most influential predictors based on permutation tests.3 Results: During the RF-based feature selection step, no predictor from the LIs of GMV, demographic or clinical variables significantly contributed to verbal fluency prediction. Instead, PET LIs of the fusiform, inferior and middle temporal gyri were identified as the best predictors for COWA (Figure 1A), and PET LIs of postcentral, inferior parietal, inferior and middle temporal gyri, and caudate were the most significant predictors for AN (Figure 1B). Successful prediction via selected PET LIs was further confirmed via leave-one-out cross-validation (Fig1CD). Using univariate models, we show that the predictors were positively correlated with verbal fluency, with more marked ipsilateral hypometabolism being associated with poorer performance (Figure 2). Conclusions: Leveraging the multivariate and nonlinear nature of RF, we identify features of brain metabolism that are collectively informative of language function in left TLE patients, and outperform underlying structure, demographic, and clinical information. Our results suggest that the disruption in baseline metabolic properties of language network areas may contribute to the underlying mechanism of language dysfunction in these patients. As yet to be tested, such metabolic deficiency may explain the disruption in brain dynamics during linguistic processing, leading to altered behavior.
References:
Balter, S et al. (2019) Brain Lang, 193:31-44
He, X et al. (2018) Brain, 1141:1375–89
Altmann, A et al. (2010) Bioinformatics, 26: 1340-7.
Funding: Please list any funding that was received in support of this abstract.: NINDS: R01-NS099348-01(DSB); R01-NS112816-01(JT)