Assessment of outcomes and machine Learning-based models to predict local failure risk following stereotactic radiosurgery for small brain metastases.

Abstract

Introduction: We assessed the outcomes of stereotactic radiosurgery (SRS) for small intact brain metastases (SBM) (≤ 2 cm) and developed machine learning (ML) algorithms to predict the probability of local failure (LF).

Methods: Consecutive patients with SBM treated with SRS between January 2017 and July 2022 were included. Propensity score matching (PSM) was performed with related factors to enhance balance for comparison. Variable selection and three time-varied generalized estimating equations (GEE) were used to create predictive models.

Results: 1503 SBMs in 235 patients treated over 358 SRS courses were analyzable. The actuarial 1-year cumulative rate of LF was lower in lesions treated with 24 Gy (5.9%, 95% CI: 4.2-8.2%) or 22 Gy (7.7%, 95% CI: 5.3-11.0%) compared to 20 Gy (25.3%, 95% CI: 18.1-34.7%) (p < 0.001). 22 Gy and 24 Gy were associated with a 63% and 74% reduction in risk in LF compared to 20 Gy (HR: 0.37; 95% CI: 0.24-0.57; p < 0.005 and HR: 0.26; 95% CI: 0.17-0.39; p < 0.005, respectively). The generated models could recommend the best dose with an individualized percentage probability of LF with each dose at 6 months, 1 year, and 2 years with a minimum AUC of 0.75. The 1-year model had the highest AUC (0.88), accuracy (88%), and specificity (91%), while the 2-year model had the highest sensitivity (89%).

Conclusion: The ML models developed predict LF as a function of dose which could aid in clinical decision-making to select an appropriate dose for SBM to optimize tumor control outcomes and schedule appropriate follow-up.