NTCP Modeling for High-Grade Temporal Radionecroses in a Large Cohort of Patients Receiving Pencil Beam Scanning Proton Therapy for Skull Base and Head and Neck Tumors.

ABSTRACT

PURPOSE: To develop a normal tissue complication probability model including clinical and dosimetric parameters for high-grade temporal lobe radionecroses (TRN) after pencil beam scanning proton therapy. METHODS AND MATERIALS We included data on 299 patients with skull base and head and neck tumors treated with pencil-beam scan proton therapy, with a total dose of ≥60 GyRBE (relative biological effectiveness) from May 2004 to November 2018. We considered 9 clinical and 27 dosimetric parameters for the structure-wise modeling of high-grade (grade ≥2) TRN. After eliminating strongly cross-correlated variables, we generated logistic regression models using least absolute shrinkage and selection operator regression. We performed bootstrapping to assess parameter selection robustness and evaluated model performance via cross-correlation by assessing the area under the curve of receiver operating characteristic curves and calibration with a Hosmer-Lemeshow test statistic. RESULTS: After a median radiologic follow-up of 51.5 months (range, 4-190), 27 patients (9%) developed grade ≥2 TRN. Eleven patients had bitemporal necrosis, resulting in 38 events in 598 temporal lobes for structure-wise analysis. During our bootstrapping analysis, we found that the highest selection frequency was for prescription dose, followed by age, V40Gy (%), hypertension, and dose to at least 1 cc (D1cc) (Gy) in the temporal lobe. During our cross-validation, we found that age*prescription-dose*D1cc (Gy)*hypertension was superior in all described test statistics. We built full cohort structure-wise and patient-wise models with maximum area under the curve of receiver operating characteristic curves of 0.79 (structure-wise) and 0.76 (patient-wise). CONCLUSIONS: While developing a logistic regression normal tissue complication probability model to predict grade ≥2 TRN, the best fit was found for the model containing age, prescription dose, D1cc (Gy), and hypertensive blood pressure as risk factors. External validation will be the next step to improve generalizability and potential introduction into clinical routine.