Abstract: Objective Bayesian Additive Regression Tree(BART)model was applied to personalize efficacy of surgery-only and surgical adjuvant chemotherapy regimens for patients with non-small cell lung cancer(NSCLC), and to identify high yield subgroups. Methods The data of all NSCLC patients from 2012 to 2013 were extracted from the SEER database. With the overall survival time no less than 5 years as the outcome, the treatment regimens were compared. The propensity scores matching method was used to equalize the covariates between the two groups, the BART model was applied to calculate the individual treatment effect(ITE), and subgroup analysis was performed by decision tree model. Results The area under the curve(AUC)of the BART model was larger than 0.8, indicating that the results had high confidence. Most patients had better efficacy with surgical adjuvant chemotherapy regimen. Important subgroup taxonomic features in clinical practice such as AJCC stage, location of primary foci and degree of tumor metastasis were obtained and the subgroup with the best efficacy was derived. Among these subgroups, compared to surgery-only options, surgical adjuvant chemotherapy regimens produced better outcomes in female patients with distant metastasis and primary site in the middle lobe of the lung. Conclusion The constructed BART model can effectively evaluate personalized efficacy of the two treatment regimens for NSCLC patients and provide a basis for personalized medical practice for similar diseases in the future.

Key words: Bayesian Additive Regression Tree, Individual treatment effect, Non-small cell lung cancer, Surgery, Adjuvant chemotherapy