Intraoperative Electrophysiologic Biomarkers Predict Critical Speech Sites in Glioma-Infiltrated Cortex.

Introduction: Diffuse gliomas invade speech-critical cortex, making maximal safe resection difficult. Resection of functional speech cortex causes deficits and reduces quality/duration of life. Direct cortical stimulation (DCS) identifies functional (DCS+) vs. nonfunctional (DCS–) cortex during awake mapping. We analyzed neuronal recordings to identify DCS+ biomarkers and trained ML prediction models to improve surgical safety.

Methods: During prospective glioma resections (2016–2024), subdural electrode arrays recorded resting-state (RS) and task (picture naming; PN) local field potentials (LFPs). Sites were classified as DCS+ vs DCS- (≥3/5 errors during stimulation). Analyses were stratified by pathology: low-grade glioma (LGG, oligodendroglioma/astrocytoma) and glioblastoma (GBM). Power in Alpha, Beta, Gamma, and High Gamma bands was compared between DCS+ and DCS– sites. Stacked logistic regression and XGBoost models were trained using 80% resting-state data and validated (20%) to classify DCS+ sites. Models were assessed via AUROC and accuracy using 10,000 permutation tests.

Results: From 91 glioma surgeries (43 LGG, 48 GBM), 1421 cortical sites were studied, with 564 sites (LGG: 29 DCS+, 278 DCS–; GBM: 20 DCS+, 237 DCS–) aligned to electrodes. PN spectrograms showed more task-related Gamma/High Gamma activation in LGG DCS+ vs DCS- sites than in GBM. At RS, LGG DCS+ sites had higher Alpha, Beta, Gamma, and High Gamma power (p < 0.05), whereas GBM DCS+ sites showed only higher High Gamma power (p < 0.05), suggesting pathology-dependent neuronal circuit differences. The LGG resting-state classifier had AUROC 0.91, sensitivity 0.93, specificity 1.0. The GBM resting-state model had AUROC 0.77, sensitivity 0.94, specificity 0.67. Classification accuracy: LGG 94%, GBM 92% (p < 0.001).

Conclusion : This study is the first to show (task and resting-state) spectral differences between DCS+ and DCS- glioma-infiltrated cortex. Further, it is the first to show resting-state biomarkers predicting DCS+ cortex in LGG and GBM. Clinically, ML-driven identification of functional speech areas via electrophysiologic features may aid in safer glioma (and other tumor) resections from eloquent cortex. Scientifically, as DCS+ sites in LGG and GBM show distinct electrophysiologic signatures, it warrants further investigation into pathology/grade-specific neural circuit effects.