1030 - Neuroanatomic Dose Constraints for Language Preservation following Brain Radiotherapy

08:05am - 08:10am ET

Room 254

Presenter(s)

Edmund Qiao, MD - University of California San Diego, San Diego, CA

E. M. Qiao¹, A. C. Puett¹, R. Karunamuni¹, J. S. Kohli², H. Nguyen¹, A. Saha¹, A. B. Hopper¹, K. R. Tringale¹, P. Sanghvi¹, V. Moiseenko¹, C. McDonald^1,2, and J. A. Hattangadi-Gluth¹; ¹Department of Radiation Medicine and Applied Sciences, University of California San Diego, La Jolla, CA, ²Department of Psychiatry, University of California San Diego, La Jolla, CA

Purpose/Objective(s): Language deficits, particularly in verbal fluency (VF), are a sequelae of brain radiotherapy (RT), yet dose constraints to neuroanatomic language substructures are not defined. We used machine learning (ML)-based normal tissue complication probability (NTCP) modeling to identify predictors of 6-month VF decline, then internally validated and optimized RT dose constraints for identified predictors.

Materials/Methods: From a prospective trial, we identified patients with primary brain tumors who received fractionated RT and had high resolution MRI/diffusion tensor imaging with verbal fluency assessments (Delis-Kaplan Executive Function System Category Fluency [DKEFS-CF] and Letter Fluency [DKEFS-LF]) pre-RT and 6 months post-RT. Automated parcellation of cortical, subcortical, and white matter association tracts from MRI was performed and neuroanatomic regions subserving language were selected and manually verified. Extreme gradient boosting modeled VF decline, defined as relative change index adjusted for practice effects (RCI-PE) =-0.67 (~25^th percentile) on both DKEFS-CF and DKEFS-LF at 6 months post-RT. Inputs included 53 demographic/clinical (e.g. age, tumor type, systemic therapy) and 168 radiation dose/neuroanatomic parameters from language-associated regions of interest. We split data 75/25 into train/test. Area under the receiver operator characteristic curve (AUC) evaluated model performance. Top 10 features (by total gain) were internally validated with univariate/multivariate logistic regressions. Cut point analysis optimized for positive predictive value (PPV) identified candidate dose/volume constraints for ML-identified regions.

Results: Among 65 patients, 16 (25%) had RCI-PE =-0.67 on DKEFS-CF and DKEFS-LF. Our model’s train/test AUCs: 0.81/0.81. The top 10 predictors: left hemisphere (LH) pars orbitalis median dose, max dose covering 2% (D02) of inferior longitudinal fasciculus, LH pars orbitalis volume (V) receiving 55 Gy (V55Gy), LH Broca superficial white matter (WM) V55Gy, LH superiortemporal WM V10Gy, LH Broca WM D02, LH pars triangularis V55Gy, left inferior frontooccipital fasciculus V35Gy, LH superiortemporal D02, and LH pars orbitalis V45. On univariate models, LH pars orbitalis V45Gy/V55Gy and LH pars triangularis V55Gy were significant (p<0.05); LH pars orbitalis V55Gy remained significant on multivariate analysis. PPV-optimized dose cut points: LH pars orbitalis V45Gy < 77% (PPV 63%), LH pars orbitalis V55Gy < 58% (PPV 80%), and LH pars triangularis V55Gy < 89% (PPV 75%).

Conclusion: Predictors of 6-month language decline after brain RT were primarily left hemisphere dosimetric features, with LH pars orbitalis V45Gy/V55Gy and LH pars triangularis V55Gy identified as significant dose-limiting substructures. These are evidence-based results for actionable cognitive-sparing, specifically for language preservation.