ESTRO 2024 - Abstract Book

S972

Clinical - CNS

ESTRO 2024

On multivariate analysis automated bootstrapped logistic regression showed the most frequently selected variable was max dose to WM underlying the left pars opercularis. Model performance by AUC was 0.80 (0.63-0.96). This was also the most frequently selected variable in LASSO (AUC 0.80, 0.64-0.95) and random forest (AUC 0.65, 0.47-0.83) modeling. The top 5 important variables in the random forest, as ranked by mean decrease in Gini coefficient, were D max to L pas opercularis WM, L caudal middle frontal WM D mean , R frontal pole WM D max , L caudal anterior cingulate cortex D mean and D max to L rostral anterior cingulate WM. These dosimetric variables outperformed all clinical variables. Performance at nested cross validation by AUC was 0.74 (logistic), 0.73 (LASSO) and 0.67 (RF). Review of univariate decision-stump analysis provides a cutoff dose for D max to the L pars opercularis WM of 54.9 Gy.

Conclusion:

We present the first NTCP model for decline in executive function based on dosimetric predictors within eloquent brain regions. Dose to several cortical and white matter ROIs significantly correlated with post-RT impairment, and these outperformed clinical variables as predictors of toxicity. The finding that the pars opercularis was the most frequently selected variable is particularly interesting. The Color-Word Interference portion of DKEFS involves both verbal/reading components as well as inhibition/switching; the pars opercularis is involved in verbal fluency/processing as well as go/no go tasks. Increasing dose to this region being predictive of cognitive decline in this domain inherently makes sense. These data can guide future cognitive-sparing strategies for partial brain RT.

Keywords: neurocognitive decline, dose constraints, NTCP