NTCP Modelling for High Grade Temporal Radio-Necrosis in a Large Cohort of Patients Receiving PBS Proton Therapy for Skull Base and Head and Neck Tumors

Abstract

Temporal lobe necrosis (TLN) is a radiation induced late toxicity of the brain tissue occurring in up to 20% of patients after high dose radiotherapy of brain tissue. While most TLN are asymptomatic, a minority develops high grade TLN that can severely impair the patient's health status. We therefore aimed to develop a normal tissue complication probability (NTCP) model including clinical and dosimetric parameters for high grade TLN after pencil beam scanning (PBS) proton therapy (PT).Data of 299 patients with skull base (n = 261) and Head and Neck (n = 38) tumors treated with PBS PT with a total dose of ≥ 60 GyRBE from 05/2004 - 11/2018 were included. Patients with a ≥ grade (G) 2 TLN (CTCAE v5.0 criteria) were considered as an event in the modelling process. A collection of 9 clinical and 27 dosimetric parameters was considered for structure wise modelling. After elimination of strongly cross-correlated variables (Spearman correlation coefficient > 0.8), logistic regression models were generated using forward stepwise selection. Bootstrapping was performed to assess parameter selection robustness. Model performance was evaluated via cross-correlation by assessing the area under the curve of receiver operating characteristic curves (AUC-ROC) as well as calibration with a Hosmer-Lemeshow (HL) test statistic. Binary cross-entropy (CE) was calculated to compare the likelihood of our data fitting the different models.After a median radiological follow-up of 51.5 months (range, 4 - 190), 27 (9 %) patients developed a ≥ G2 TRN. With 11 (41%) patients having bitemporal necrosis, this resulted in 38 events in 598 temporal lobes for structure-wise analysis. After multicollinearity parameter removal, 7 clinical and 7 dosimetrical parameters were used for further analysis. During the Bootstrapping analysis the highest selection frequency was found for prescription dose (PD), followed by Age, Diabetes (DM), Hypertension (HBP) and D1ccGy and DminGy. Changes in Bayes information criterion were minimal for models > 3 parameters. As a result, models with 3 and 4 parameters were chosen for further evaluation. DM and DminGy failed to reach significance in the generated models, so models including HBP, PD, age and D1ccGy were selected for performance evaluation. During cross validation, Age*PD*D1ccGy and Age*PD*D1ccGy*HBP were superior in all described test statistics. Full cohort structure wise and patient wise models were built with a maximum AUC-ROC of 0.7931 (structure wise) and 0.7590 (patient wise) for the model including Age*PD*D1ccGy*HBP.While developing logistic regression NTCP models to predict ≥ G2 TLN, the best fit was found for the model containing Age, prescription dose, D1ccGy of the temporal lobe and high blood pressure as risk factors. External validation would be the next step to improve generalizability and potential introduction into clinical routine, allowing for patient-specific planning to avoid high grade TLN for high-risk patients.C. Schroeder: None. A. Köthe: None. C. De Angelis: None. L. Basler: None. D. Leiser: None. A. Lomax: None. D.C. Weber: None.