Local control and dose selection for lung cancer brain metastases treated with radiosurgery: an artificial intelligence model

Clinical Snapshot

Key Findings

• P Value: Not reported

• Effect Size: C-index 0.83 for survival prediction model

• Primary Outcome: Local control prediction accuracy

• Nnt Or Sensitivity: Integrated Brier Score 0.15, baseline dose prediction MAE 1.59 Gy

• Confidence Interval: Not reported

Clinical Application

Requires sophisticated computational infrastructure and access to comprehensive clinical and dosimetric data Relevant to Australian neurosurgical centers offering Gamma Knife or LINAC-based SRS; could inform PBS funding decisions for optimal dose selection and reduce treatment variability Patients with lung cancer brain metastases suitable for stereotactic radiosurgery

Abstract

Clinical and radiological outcomes after Stereotactic radiosurgery (SRS) for lung cancer brain metastases are heterogeneous, and prescription dose selection is often guided by experience rather than individual patient and tumor specific features. A data-driven approach that links dose to quantitative local control and adverse effect predictions could improve planning and follow-up strategies. We performed a retrospective single-center cohort study of lung cancer brain metastases treated with GKRS at the University of Pittsburgh Medical Center (2014-2024). Using routinely available clinical, tumor, and dosimetric variables, we trained a tumor-level Random Survival Forest to predict Local Control and Dose Selection for Lung Cancer Brain Metastases after GKRS treatment. Internal validation used leakage-resistant patient-level grouped cross-validation. Performance was assessed using Harrell's concordance index (C-index) and integrated Brier score (IBS). A connected dose-sweep decision layer evaluated predicted local control across a clinically feasible margin dose grid and selected the dose associated with the most favorable predicted profile at prespecified horizons. The survival model demonstrated good internal performance (C-index 0.83; IBS 0.15). A baseline dose imitation model predicted historical margin dose with low error (OOF MAE 1.59 Gy). The integrated decision layer generated individualized local control trajectories and returned horizon-specific local control probabilities and model-based prescription dose recommendations. An Artificial Intelligence framework integrating time-to-event prediction with outcome-linked dose sweeping can provide individualized GKRS decision support for lung cancer brain metastases by delivering quantitative local control forecasts and model-based dose recommendations.

References

1. 1.Reyes, J. S., Bouras, A., Hadjipanayis, C. G., Niranjan, A., & Lunsford, L. D. (2026). Local control and dose selection for lung cancer brain metastases treated with radiosurgery: an artificial intelligence model. Clinical & Experimental Metastasis. https://doi.org/10.1007/s10143-025-03597-9