Personalized prediction of local control after Stereotactic radiosurgery for craniopharyngioma: a multicenter machine learning Survival model

Abstract

Stereotactic radiosurgery (SRS) is used in selected patients with craniopharyngioma, yet counseling and follow-up planning often rely on population-level local control rates rather than individualized expectations over time. Objective To develop and internally validate a multicenter survival model to predict imaging-defined time to progression after SRS for craniopharyngioma. Methods We analyzed a multicenter IRRF registry of SRS-treated craniopharyngioma patients. Imaging progression was defined by the overall last imaging response (PD vs. non-PD), with censoring at last imaging follow-up when progression was not observed. A Random Survival Forest (RSF) model was evaluated using 5-fold out-of-fold cross-validation. Performance was assessed using the concordance index, time-dependent AUC at 12, 24, and 60 months with bootstrap 95% confidence intervals, integrated Brier score (IBS) over 0–60 months, and risk-stratified calibration. Benchmarks included a penalized Cox model and a Kaplan–Meier baseline. Results Among 277 patients (event rate 13.0%; median imaging follow-up 57.0 months by reverse Kaplan–Meier), RSF achieved an out-of-fold C-index of 0.905. Time-dependent AUC was 0.895 (95% CI 0.828–0.959) at 12 months, 0.897 (95% CI 0.833–0.952) at 24 months, and 0.934 (95% CI 0.889–0.969) at 60 months. IBS (0–60 months) was 0.050 with favorable calibration. Conclusions A multicenter machine learning survival model can provide individualized, well-calibrated estimates of local control over time after SRS for craniopharyngioma to support non-prescriptive decision support. Clinical trial number Not applicable.