Re-evaluating the neutrophil-to-lymphocyte ratio: Machine learning-based variable selection for predicting survival at twelve months in late-stage cancer patients receiving immunotherapy.

Abstract

e18201 Background: Neutrophil to Lymphocyte Ratio (NLR) is prognostic for cancer patients treated with immune checkpoint inhibitors (ICI). We showed the change in NLR early during treatment to be a stronger, curvilinear predictor, i.e. patients with an intermediate change in NLR performed better than those with large decreases or increases. This led us to re-examine whether NLR is an optimal predictor of overall survival (OS). Methods: A retrospective review of 467 patients with advanced cancer who received ICIs from 2011 to 2017 at the Ohio State University was performed with IRB approval. NLR was collected at the initiation of ICI and on-treatment (median 21, IQR 8 days) and calculated as ratio of absolute neutrophil to lymphocyte counts. Variable selection machine-learning algorithms included fast and frugal decision trees and random forest, performed in R. Results: The machine-learning algorithm fast and frugal decision trees identified the ratio of NLR on treatment to baseline NLR, the NLR on treatment, the change in NLR and the cubic change in NLR to be the most informative predictors of survival at 12 months. A random forest algorithm identified the same four variables as the most important for prediction accuracy. Age, sex and cancer type were the least informative predictors in the model, suggesting the on-treatment NLR variables are of value across wide range of demographics. Conclusions: NLR measured during treatment, and its derivative values of the ratio to baseline, change from baseline, and the cubic change from baseline, hold more predictive value than NLR measured at baseline. Common control variables such as age, and sex showed little effect on the model, suggesting on-treatment NLR is useful across wide demographic space. [Table: see text]