NCMP-06. A machine learning predictive model of aphasia in patients with diffuse glioma using radiomic, molecular, and clinical features

Gliomas infiltrate eloquent cortex yet continue to participate in language processing, albeit with reduced information encoding. The molecular determinants of this dysfunction are unknown. The present work aims to define imaging and genomic predictors of aphasia in diffuse supratentorial gliomas. We studied 88 adults (median 59 y; 38 % female) with newly diagnosed WHO grade 2-4 gliomas who underwent targeted next-generation sequencing and Quick Aphasia Battery (QAB) testing (2017-2021). Support-vector regression lesion–symptom mapping localized voxels where tumor burden predicted QAB scores. A random-forest model integrating radiomics, molecular variants, and clinical variables ranked predictors of aphasia; multivariable linear regression validated key features. Single-nucleus RNA-seq of resected tumors clarified cellular expression patterns. Lesion–symptom mapping isolated a 61,745-voxel cluster in the left superior temporal gyrus whose tumor overlap strongly correlated with lower QAB scores. The random-forest model highlighted three radiographic metrics—T1-post-gadolinium, T2-FLAIR, and total lesion voxels within this cluster—and seven genomic alterations. The top molecular predictor, ARID2 mutation (3/88), was associated with markedly worse language performance (mean ± SD QAB 6.6 ± 1.4 vs 9.1 ± 1.3; p = 0.01). On multivariate analysis, ARID2 status (β = –1.42; 95 % CI –2.63 to –0.21; p = 0.02) and contrast-enhancing voxels in the cluster (β = –0.08 per 100-voxel increase; p < 0.001) independently predicted aphasia. Single-nucleus RNA-seq revealed ARID2 expression enriched in proliferating tumor cells, linking chromatin-remodeling dysregulation to cortical network dysfunction. An imaging-genomic signature comprising ARID2 mutation and contrast-enhancing burden in a dominant temporal language hub robustly predicts aphasia in diffuse glioma. These findings nominate ARID2 as a mechanistic driver of glioma-associated language impairment and provide actionable markers for precision surgical and adjuvant strategies to preserve cognition