Background: In recent years, there have been expectations for the application of artificial intelligence (AI) in the clinical practice of lung cancer. In the first-line setting, immune checkpoint inhibitors (ICIs) in combination with cytotoxic agents (CTx) have become a standard-of-care for patients with non-small cell lung cancer (NSCLC); however, no absolute biomarkers have been established for the prediction of the response to such therapy. In this study, we aimed to construct a new prognostic model by using AI in NSCLC patients treated with first-line ICI plus CTx. Methods: We retrospectively identified 218 NSCLC patients with advanced or postoperatively recurrent disease. They received first-line treatment with ICI alone or in combination with cytotoxic agents between January 2016 and October 2020. AI model using Gradient Boosting Decision Tree (GBDT) was used to identify predictors (explanatory variables) of the survivals (objective variables). A total of 53 factors were selected for the explanatory variables including clinical, hematological and pathological information. The objective variables were overall survival (OS) and progression-free survival (PFS). GBDT constructed the optimal predicting model, and calculated the contribution of each explanatory variable to the objective variables. Additionally, the importance of each feature was interpreted by SHapley Additive exPlanation (SHAP). Results: The median follow-up period was 14.4 months (range: 0.5-52.0) after the initiation of the therapy. The characteristics were shown below; the median age was 69 years old (range: 36-85); males and females were 165 (75.7%) and 53 (24.3%); patients with ICI alone or combination with cytotoxic anticancer agents were 91 (41.7%) or 127 (58.3%); the distributions of performance status (PS) were 107 (49.1%), 94 (43.1%), 13 (6.0%) with PS 0, 1 and 2; the Sq, non-Sq and others were found in 51 (23.4%), 165 (75.7%) and 2 (0.9%); 170 patients (78.0%) showed a PD-L1 expression of > 1% and 36 (16.5%) patients had no PD-L1 expression. The median PFS and OS were 10.2 and 21.8 months, respectively. The predicting optimal model for OS and PFS were constructed by GBDT, and the respective area under the curves (AUC) were 0.724 and 0.656. The explanatory variables indicating high contribution to OS were serum C-reactive protein (CRP), cytokeratin subunit 19 fragment (CYFRA), alkaline phosphatase (ALP), peripheral lymphocyte count and chloride ion level. As for PFS, the highest five contributions were CYFRA, CRP, immune-related adverse events, ALP and peripheral platelet count. Furthermore, all of these factors were explained their importance by SHAP. Neither PD-L1 expression nor PS were predictors of the survivals. Conclusions: Machine learning identified several novel factors possibly contributing to the prognoses in NSCLCs patients treated with the first-line immunotherapy.