Development of clinically accessible nomograms to predict risk of brain metastases at baseline and follow-up in patients with non-small cell lung cancer.

2035 Background: Brain metastases (BM) are a common complication in non-small cell lung cancer (NSCLC). Reliable models predicting risk of BM development are lacking, hindering effective CNS screening and patient prognostication. In the era of precision medicine, these are important gaps in our knowledge. The aims of this study were to 1) evaluate published BM risk-stratification algorithms, and 2) develop nomograms to predict BM incidence. Methods: Using a retrospective cohort of NSCLC patients from Penn State Health (2011-2020), we 1) evaluated the performance of published BM risk-stratification algorithms systematically identified, and 2) developed nomograms to predict risk of BM incidence. For Aim 1, published algorithms were benchmarked using AUROCs calculated from logistic regression models. For Aim 2, cox-proportional hazard models were trained using L1-regularization, and nomograms were constructed to predict BM risk at 6-month, 1-year, and 2-year follow up. Two separate nomograms were developed: Model T0 used only clinical and imaging data available at time of diagnosis, while Model T1 leveraged additional molecular characteristics and treatment history. All models were trained using 70% of data and tested using 30% of data. Time-dependent AUROC metrics for nomograms were calculated using a cumulative sensitivity and dynamic specificity-based estimator. Results: Our cohort included 1904 patients (median age 68, range: 38 to 94 years, BM incidence 22.8%). Aim 1: 12 published algorithms were identified that used variables consistently available in patient charts. Among these, the Zhang 2021 model was the best predictor of cumulative BM risk (AUROC 95% CI = 0.89 0.85-0.93). Aim 2: Model T0 was trained using age at diagnosis and clinical TNM stage and predicted BM incidence at 6-month, 1-year and 2-year follow up with AUROCs of 0.87, 0.85, and 0.87, respectively. Model T1 was trained with additional predictors, including number of extra-cranial metastatic sites, treatment history (e.g., radiation, surgery, chemotherapy, etc.), and mutation profile (EGFR, KRAS, ALK, BRAF), and achieved AUROCs of 0.90, 0.89, and 0.91 at 6-month, 1-year and 2-year follow up, respectively. Distant metastases at time of NSCLC diagnosis (HR 95% CI = 3.38 2.28, 4.99) and number of extra-cranial metastatic sites (HR 95% CI = 1.75 1.54, 1.99 per each additional metastasis) were the strongest independent predictors of BM risk. Conclusions: Based on one of the largest NSCLC cohorts to date, we have developed clinically accessible nomograms for prediction of BM development. This tool can be readily applied toward prognostic modeling and risk stratification, refinement of practice guidelines for CNS screening, and patient counseling.