A Brain Metastases Survival Model Using an Ensemble Tree Approach

Abstract

<h3>Purpose/Objective(s)</h3> The primary purpose of this study is to determine whether a machine learning approach can estimate survival in patients with brain metastases undergoing stereotactic radiosurgery or fractionated stereotactic radiotherapy (SRS/SRT). The secondary purpose is to identify variables of importance. <h3>Materials/Methods</h3> Data were collected for 285 SRS/SRT treatments in 228 patients. If a patient was treated with more than one course of SRS/SRT within 30 days, they were counted only once. Twenty-five clinically-relevant variables were identified as covariates and the primary outcome of time from brain metastasis diagnoses to death was used to build a random survival forest model. Brain metastasis location was categorized as infratentorial, supratentorial, or both. An 80/20 split was used for training (n = 228) and test (n = 57) sets. Missing data points were imputed using a just-in-time adaptive tree approach. Minimal depth and variable importance (VIMP) approaches were used to identify prognostic factors. Model performance was assessed using time-dependent area under the receiver operating characteristics curve (tAUC). <h3>Results</h3> Median survival time was 17 months. The most important variables according to minimal depth analysis (depth threshold 5.49) were age, extracranial status, Karnofsky Performance Status (KPS), intracranial metastases volume, primary cancer histology, insurance status, number of intracranial metastases, SRS/SRT dose, molecular marker status, and brain metastasis location (see table). Error rate on the test set was 0.27. tAUC was found to increase continuously over time and at 3, 6, 12, and 24 months was 0.62, 0.72, 0.77, and 0.86 respectively. <h3>Conclusion</h3> An ensemble tree approach can provide good survival prediction for patients with brain metastases undergoing SRS/SRT. Model performance, as measured by tAUC, increases over time suggesting better predictive capability at longer time intervals. Future directions include collecting more data to increase model performance, comparing to other models, and validating with external data.