Clinical GBM hybrid artificial intelligence for prescription dose recommendation and outcome prediction after gamma knife radiosurgery treatment: a proof-of-concept

Source abstract

Background Outcomes after Gamma Knife radiosurgery (GKRS) for recurrent glioblastoma (GBM) are heterogeneous, and prescription dose selection remains challenging in previously irradiated brain. A data-driven approach that individualizes dose while providing quantitative, patient-specific outcome predictions could improve counseling and follow-up planning. Methods We performed a retrospective single-center study of recurrent GBM treated with GKRS at the University of Pittsburgh Medical Center (2014–2024) to develop CGH-AI (Clinical GBM Hybrid Artificial Intelligence). The patient-related clinical variables, biopsy-derived markers, and dosimetric parameters were extracted from the medical records. The primary endpoint was local tumor control, modeled as time to local failure. A lesion-level survival model was developed using a Random Survival Forest and internally validated using patient-level grouped cross-validation. Discrimination and calibration were assessed using Harrell’s concordance index (C-index) and integrated Brier score (IBS). A connected dose recommendation engine evaluated clinically feasible candidate doses per case and selected the dose associated with the most favorable predicted local control profile. Results The local control survival model achieved strong patient-level internal validation performance (C-index 0.80; IBS 0.14). Sensitivity analyses incorporating biopsy-derived markers yielded similar performance, supporting robustness of the clinical–tumor–dosimetric feature set. The integrated dose recommendation engine consistently identified prescription doses associated with improved predicted local control while providing interpretable, case-specific local control probabilities and expected local control duration. Conclusions CGH-AI integrates survival modeling and dose decision support within the GKRS workflow to recommend individualized prescription dose and to predict local control for recurrent GBM, supporting personalized decision-making and surveillance planning.