2645 - Development of a Machine Learning-Based Tool to Predict Local Failure after Stereotactic Radiosurgery for Small Brain Metastases

S. Yarlagadda¹, Y. Zhang², A. Saxena², T. Kutuk¹, R. P. Tolakanahalli^1,3, H. Appel¹, A. La Rosa¹, M. D. Hall^1,3, R. H. Press¹, D. J. Wieczorek^1,3, Y. C. Lee^1,3, T. Bejarano¹, M. W. McDermott⁴, A. Gutierrez¹, M. P. Mehta^1,3, and R. Kotecha^1,5; ¹Department of Radiation Oncology, Miami Cancer Institute, Baptist Health South Florida, Miami, FL, ²Department of Biostatistics, Miami Cancer Institute, Baptist Health South Florida, Miami, FL, ³Department of Radiation Oncology, Florida International University, Herbert Wertheim College of Medicine, Miami, FL, ⁴Department of Neurosurgery, Miami Neuroscience Institute, Baptist Health South Florida, Miami, FL, ⁵Florida International University, Herbert Wertheim College of Medicine, Miami, FL

Purpose/Objective(s): Single-fraction stereotactic radiosurgery (SRS) is widely used for the management of small brain metastases (SBM; = 2 cm). However, there is considerable variability in the prescription doses across institutions and in clinical practice guidelines. With the recent advances in artificial intelligence, we aimed to develop a machine learning (ML) algorithm to model the relevant patient-, disease-, and treatment-related factors to predict the probability of local failure (LF) as a function of dose. Materials/

Methods: Consecutive patients with intact SBM treated with SRS between January 2017 and July 2022 were included. Patient baseline characteristics and treatment-related parameters were extracted from electronic medical records. To limit the integral brain dose when treating multiple brain metastases, the institutional practice was to reduce the prescription dose with an increase in the number of lesions treated (i.e. 24 Gy for =10 lesions, 22 Gy for 11-20, 20 Gy for >20). Propensity score match (PSM) analysis was performed with factors such as age, sex, race, Karnofsky performance score (KPS), maximum lesion size, total number of lesions, extracranial disease burden and status, course of SRS, systemic therapy type, and adjusted before creating the model. Considering this clustered data structure, generalized estimating equations (GEE) with a threshold of 0.5 was used to create a supervised machine-learning model, an extension of a generalized linear model with binomial distribution for binary classification to identify the probability of LF with each prescription dose. The data were randomly divided (70% training set, 30% test set) to develop and evaluate the model.
Results: 1503 SBMs in 235 patients that received SRS were analyzable. The median age was 65 years (Interquartile range [IQR]: 55-73) and majority were female (61%). Median KPS was 90 (IQR: 80-90), median number of lesions per SRS course was 4 (IQR: 2-7). The most common primary tumors were lung (58.5%), followed by breast cancer (24.6%). The prescription dose was 20 Gy for 297 lesions (20%), 22 Gy for 442 lesions (29%), 24 Gy for 764 lesions (51%). With a median follow-up of 12 months (IQR: 4-23), LF occurred in 138 lesions (9.2%) in 47 patients. After propensity score matching, 276 lesions (123 patients) were included in the GEE model. The model included four variables- prescription dose, age, KPS, and first course SRS. The model could recommend the best dose for each patient with an individualized percentage probability of LF with each dose with a test accuracy of 83%. F1 score of identifying LF was 0.74, AUROC 0.78, precision 0.83, and recall 0.67.
Conclusion: In the patient cohort, the ML model developed in this study was able to predict LF as a function of dose with an accuracy of 83% and identified key parameters associated with LF, including dose, age, KPS, and course of SRS. This algorithm could aid in clinical decision-making to select an appropriate dose for SBM to optimize tumor control outcomes.