BIOS-01. MACHINE LEARNING-BASED PREDICTION OF INPATIENT CARE COSTS FOR PATIENTS WITH BRAIN METASTASES: A REAL-WORLD ANALYSIS

Abstract INTRODUCTION Brain metastases (BMs) are increasingly a cause of significant cancer-related morbidity, mortality, and economic burden. While BM-related healthcare costs have been previously described, predictive analytics based on contemporary cohorts, particularly for inpatient care, remain unavailable. This study sought to evaluate inpatient care costs (IPCC) and associated reimbursements for patients with BMs and perform predictive modeling through supervised machine learning (sML) using a contemporary, commercial US claims database. METHODS This retrospective study, reported following STROBE and TRIPOD+AI reporting guidelines, included adult patients (aged 18-65 years) in the United States (US) with BMs included in the MarketScan Commercial Claims and Encounters Database who received inpatient care during January 2016 to December 2021. Eligible patients were identified using ICD-10 codes. IPCC measures were adjusted using the Medical Care component of the 2023 Consumer Price Index. The primary study outcome was total IPCC, represented by adjusted gross total payments. After log-transformation performed due to data skew, sML was carried out using Random Forest algorithm with 1000 iterations with a 70:30 training-test data split, followed by residual analysis and evaluation of predictive performance using root-mean-squared error (RMSE), out-of-bag error (OOBE), and calibration curve. Multivariable linear regression was carried out for sensitivity analyses. RESULTS A total of 25117 unique inpatient admissions across 13477 BM patients were included, with median (IQR) age of 57 years (51-61). Gross median (IQR) CPI-medical-care-adjusted values per admission were: total payments $28,119.51 (16847.45-55111.67); hospital payments $24,678.74 (14601.66-47864.84); and payments to principal physician $884.80 (441.07-2434.15), being 3.6% (IQR 1.9-6.8) of total payment. Median inpatient length of stay was 4 days (IQR 2-8). ML model had an OOBE of 0.61 and, using test data, RMSE of 0.85, indicating robust performance. No major deviations were found on residual analysis. CONCLUSIONS Contemporary IPCCs for BM patients in US are substantial but driven minimally by payments to principal physician. sML may be utilized to successfully predict IPCC for BM patients.