NCOG-05. A BRAIN METASTASES SURVIVAL MODEL USING AN ENSEMBLE TREE APPROACH

Abstract PURPOSE/OBJECTIVE(S) 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 radiotherapy (SRS/SRT). secondary identify covariates importance. MATERIALS/METHODS Data were collected for 377 SRS/SRT treatments 291 done between the years 2008-2021. If patient was treated more than one course within 30 days, they counted only once. Twenty-five clinically-relevant variables identified as and outcome time from metastasis diagnoses death used build random forest model. Brain location categorized infratentorial, supratentorial, both. An 80/20 split training (n = 302) test 75) sets. Missing data points imputed using just-in-time adaptive tree approach. Minimal depth variable importance (VIMP) approaches prognostic factors. Model performance assessed time-dependent area under receiver operating characteristics curve (tAUC). RESULTS Median 16 months. most important according minimal analysis (depth threshold 5.23) Karnofsky Performance Status (KPS), extracranial status, age, insurance volume, histology, number metastases, location. Error rate on set 0.38. tAUC found increase continuously over at 6, 12, 24, 36 months 0.56, 0.63, 0.74, 0.84 respectively. CONCLUSION ensemble provide good prediction SRS/SRT. performance, measured by tAUC, increases suggesting better predictive capability longer intervals. Future directions include collecting model comparing other models, validating external data.