Evidence-Based Multivariate Normal Tissue Complication Probability (NTCP) Study of Domain-Specific Cognitive Decline after Partial Brain RT

Abstract

Beyond the hippocampus, there are no evidence-based dose constraints for eloquent brain structures which subserve memory and attention/processing speed. We performed a multivariate normal tissue complication probability analysis of post-RT neurocognitive decline, examining dosimetric predictors of eloquent brain regions. Data were analyzed from a prospective longitudinal clinical trial. Patients (n = 78) with primary brain tumors receiving fractionated RT complete a comprehensive neurocognitive evaluation and high-resolution volumetric and diffusion MRI at baseline and 6 months post-RT. Image processing using robust, validated automated segmentation parcellated individual WM tracts, cortical regions, and hippocampi. Well-validated neurocognitive tests including Delis-Kaplan Executive Function System and Wechsler Adult Intelligence Scale-IV coding (attention/processing), Boston Naming Test (language) and Hopkins Verbal Learning Test and Brief Visuospatial Memory Test (verbal/visuospatial memory) were assessed. Reliable change indices adjusted for practice effects (RCI-PE) were calculated for each patient between baseline and 6 months; a negative RCI-PE was scored as decline. Univariate logistic regression was performed with mean and max dose to structures of interest as well as clinical variables. Multivariate model building was performed using automated bootstrapped logistic regression, LASSO and random forest modeling. On univariate analysis mean and max dose to multiple regions of the corpus callosum (CC) were correlated with attention/processing speed decline; most significantly in WAIS coding, including Dmax to the anterior CC (p = 0.011) and central CC (p = 0.010), and Dmax and Dmean to the mid anterior CC (p = 0.006 and 0.010). Mean dose to the left fornix was associated with decline in memory (p = 0.023, cutoff 12.9 Gy), as were increasing age and both concurrent and adjuvant chemotherapy. On multivariate analysis for attention, automated bootstrapped logistic regression showed the most frequently selected variable was mean dose to the mid anterior CC. Performance at nested cross-validation by AUC was 0.80 (0.75-0.84); LASSO model performance by AUC was 0.76 (0.72-0.81) with Dmean to the mid anterior CC being the most frequent variable. The top five most important variables in the Random Forest as ranked by mean decrease in Gini coefficient were mean dose to mid anterior CC, all white matter, combined CC and max dose to CC and posterior CC. Model performance by AUC was 0.66 (0.60-0.71). Here, we present the first, to our knowledge, NTCP model for decline in attention/processing speed, along with dosimetric predictors of memory decline beyond the hippocampus. We found that after partial brain RT, dose to several ROIs significantly correlated with post-RT impairment. These data can guide future cognitive-sparing strategies for brain RT.