P01.02.A WHOLE-BRAIN STRUCTURAL CONNECTIVITY PREDICTS COGNITIVE OUTCOMES IN PRETREATED PATIENTS WITH WHO GRADE 3 OR 4 GLIOMAS

Abstract

Abstract BACKGROUND Glioma patients frequently suffer from cognitive dysfunction potentially related to brain damage caused by tumor invasion or therapeutic intervention. We hypothesized that, as in other neurological disorders, the long-term outcome of cognitive functions in glioma patients critically depends on maintaining the structural connectivity of multiple interconnected networks. PATIENTS AND METHODS One hundred and twenty-one glioma patients (median age, 52 years; median ECOG score, 1) with histomolecularly characterized CNS WHO grade 3 or 4 gliomas were investigated after a median time of 14 months (range, 1-214 months) following first-line therapy, including resection and chemoradiation with alkylating agents. Hybrid amino acid PET/MR imaging using the tracer O-(2-[18F]fluoroethyl)-L-tyrosine, a brain-networks-based cortical parcellation into nodes, and advanced fiber tractography tools were used for constructing fiber-count-weighted connectivity matrices. Cognitive performance was measured by ten different tests concerning various cognitive domains. The connectivity matrices and the patients' test scores were applied to a machine learning-based model that identified critical fiber connections, cortical nodes, and involved networks for predicting cognitive performance in the assessed domains. RESULTS Compared to a matched cohort of healthy subjects, the cognitive scores of the patients were significantly lower in 9 of 10 cognitive tests. The most affected domains were executive function/concept-shifting (-72%), attention/processing speed (relative performance, -53%), followed by semantic word fluency (-25%), verbal working and semantic memory (-8 and -22%, respectively), and visual working memory (-11 to -20%). For each cognitive test, a subset of connections (median number, 254; range, 32-542) was identified whose fiber count sum showed a correlation with the actual scores in a linear model (median R2, 0.37; range, 0.16-0.44). Leave-one-out cross-validation confirmed the model's generalizability in 7 of 10 tests (median correlation coefficient for predicted compared to observed scores, 0.47; range 0.39-0.57). The distribution of critical cortical regions (node degree of ≥ 5 predictive edges), found in 90% of the cross-validation iterations, varied considerably between domains. These included predominantly left-sided cortical nodes of the visual, somatomotor, dorsal/ventral attention, and default mode networks. Highly critical nodes (degree of 15-20 predictive edges) included the default-mode network’s left temporal and bilateral posterior cingulate cortex. CONCLUSION These results suggest that the cognitive performance of pretreated glioma patients is strongly related to the structural connectivity within multiple brain networks and depends on the integrity of known network hubs.