2032 Background: 30-50% of patients with non-early NSCLC will eventually develop BM, with a median survival of less than one year from BM diagnosis. There are no widely accepted clinical risk models for development of BM in patients without them at baseline. We predicted the binary risk of BM using clinical and genetic factors from a large multi-institutional cohort. Methods: Stage II-IV NSCLC patients from the AACR Project GENIE Biopharma Consortium dataset were eligible. This consisted of 4 academic institutions who curated clinical data of patients who had somatic next-generation tumor sequencing (NGS) between 2015-2017. We excluded patients who had BM at baseline, died within 30 days of NSCLC diagnosis, or did not undergo brain imaging. Covariates included demographics, anticancer therapies (received up to 90 days prior to BM development and within 5 years from NSCLC diagnosis), and NGS data; radiotherapy (RT) data were not available. NGS features included mutations and copy number alterations. These features were restricted to those classified as oncogenic by OncoKB. Univariate feature selection with Fisher’s test (p<.1) was performed on medication and genetic features. We compared 5 different machine learning models for prediction: random forest (RF), support vector machine (SVM), lasso regression, ridge regression, and an ensemble classifier. We split our data into training and test sets. 10-fold cross-validation was done on the training set for parameter tuning. The area under the receiver-operating curve (AUC) is reported on the test set. Results: 956 patients were included, 192 (20%) in the test set. Univariate features associated with BM were treatment with etoposide, Asian race, presence of bone metastases at NSCLC diagnosis, mutations in TP53 and EGFR, amplifications of ERBB2 and EGFR, and deletions of RB1, CDKN2A and CDKN2B. Univariate features inversely associated with BM were older age, treatment with nivolumab, vinorelbine, alectinib, pembrolizumab, atezolizumab, and gemcitabine, as well as mutations in NOTCH1 and KRAS. Ridge regression had the best AUC, 0.73 (Table). Conclusions: We achieved reasonable prediction performance using commonly obtained clinical and genomic information in non-early NSCLC. The biologic role of the associated alterations deserves further scrutiny; this study replicates similar findings for EGFR and KRAS in a much smaller cohort. Certain subsets of NSCLC patients may benefit from increased surveillance for BM and transition to drug therapies known to effectively cross the blood-brain barrier, e.g., nivolumab and alectinib. Inclusion of additional covariates, e.g., brain RT, may further improve model performance.[Table: see text]
Influenza Prediction: Analyzing Machine Learning Algorithms
