CNSC-52. SPEECH-DERIVED LINGUISTIC AND SPECTRAL FEATURES PREDICT MOLECULAR SUBTYPES OF ADULT-TYPE DIFFUSE GLIOMAS

Abstract INTRODUCTION Gliomas engage in synaptic integration with surrounding neural circuitry. Previous work demonstrated that glioma-infiltrated cortex retains the capacity to encode speech, though it relies on a more diffuse network. While behavioral variability across glioma subtypes has been reported, no prior study has evaluated whether speech characteristics can predict glioma subtype. Here, we conduct a comprehensive linguistic and spectral analysis of speech in patients with gliomas to determine whether preoperative features can predict molecular features including IDH mutation status. METHODS This study utilized a prospective registry of patients with gliomas who completed semantic naming tasks. Recordings were annotated for error types by a single blinded observer. Audio signals were filtered using notch, Butterworth, and low-pass filters before spectral decomposition. Power band and power spectral density (PSD) features were extracted. Non-parametric tests and Spearman correlations were used to compare IDH wild-type and mutant groups. A gradient boosting machine learning classifier was trained using a 70/30 train-test split. RESULTS A total of 100 patients were included. IDH wild-type patients were older (62.0 vs. 39.8 years, p<0.001), more often white (89.7% vs. 68.4%, p=0.007), and exhibited greater FLAIR volume, higher Ki67 indices, and lower MGMT promoter methylation. IDH wild-type patients exhibited higher anomia and circumlocution rates on picture naming (p = 0.018, p = 0.043) and increased overall errors, anomia, verbal paraphasias, and delays on auditory naming (p < 0.001–0.02). Spectral analysis revealed group differences in clarity and brilliance bands (p=0.029, 0.01), and PSD analysis revealed differences in bass, midrange, and brilliance power (p=0.041–0.05). The gradient boosting model achieved an AUC of 0.89, with top predictive features including age, brilliance band power, midrange PSD, and auditory naming error rate. CONCLUSIONS Pre-operative speech features differ significantly by IDH status. Spectral-linguistic profiling coupled with machine learning can non-invasively predict glioma molecular subtype.