Advances in radiation oncology, Volume 10, Issue 10, 24 4 2025, Pages 101871 Predicting Radionecrosis After Stereotactic Radiation Therapy for Solitary Brain Metastases: External Validation of a Univariable Model and Development of a Multivariable Model. van Schaik EEMW, Crouzen JA, Mast ME, Broekman MLD, Hakstege MP, Mertens BJA, Kiderlen M, van der Voort van Zyp NCMG, Petoukhova AL, Zindler JD
Purpose
Stereotactic radiation therapy (SRT) is a frequently used and effective treatment for patients who received diagnosis for brain metastases (BMs). Radionecrosis (RN) is a severe complication of SRT which may cause neurologic symptoms. A normal tissue complication probability (NTCP) model has previously been established to predict the risk of RN based on the volume of healthy brain receiving ≥12 Gy. The aim of this study is to externally validate this prediction model.
Methods and materials
A total of 162 patients treated with SRT for solitary BMs were retrospectively included. The NTCP models for all (asymptomatic and symptomatic) RN and symptomatic RN cases were evaluated using discrimination (C-statistic) and calibration (Brier scores). Overall survival was determined using the Kaplan-Meier method.
Results
Median overall survival was 10 months. Asymptomatic or symptomatic RN was found in 44 (27%) of patients. Of these, 26 (16%) RN cases were symptomatic, with actuarial rates of 11% and 22% after 6 and 12 months, respectively. The C-statistics of the "any RN" and the "symptomatic RN" models were identical (0.61). Brier scores were 0.201 and 0.217, respectively. Univariable logistic regression analysis showed a significant correlation between both tumor volume and volume of healthy brain receiving ≥12 Gy with symptomatic RN. These effects did not hold up in the multivariable analysis. A nomogram was established and internally validated.
Conclusions
The accuracy of the tested NTCP models in assessing the risk of RN in patients with BM after SRT was insufficient for clinical practice. A novel multifactorial nomogram was developed to predict symptomatic RN. This model needs to be externally validated.
