Background: Brain metastasis, with an incidence of more than 30%, is a common complication of non-small cell lung cancer (NSCLC). Therefore, there is an urgent need for an assessment method that can effectively predict brain metastases in NSCLC and help understand its mechanism. Materials and methods: GSE30219, GSE31210, GSE37745, and GSE50081 datasets were downloaded from the GEO database and integrated into a dataset (GSE). The integrated dataset was divided into the training and test datasets. TCGA-NSCLC dataset was regarded as an independent verification dataset. Here, the limma R package was used to identify the differentially expression genes (DEGs). Importantly, the RiskScore model was constructed using univariate Cox regression analysis and least absolute shrinkage and selection operator (LASSO) analysis. Moreover, we explored in detail the tumor mutational signature, immune signature, and sensitivity to treatment of brain metastases in NSCLC. Finally, a nomogram was built using the rms package. Results: First, 472 DEGs associated with brain metastases in NSCLC were obtained, which were closely associated with cancer-associated pathways. Interestingly, a RiskScore model was constructed using 11 genes from 472 DEGs, and the robustness was confirmed in GSE test, entire GSE, and TCGA datasets. Samples in the low RiskScore group had a higher gene mutation score and lower immunoinfiltration status. Moreover, we found that the patients in the low RiskScore group were more sensitive to the four chemotherapy drugs. In addition, the predictive nomogram model was able to effectively predict the outcome of patients through appropriate RiskScore stratification. Conclusion: The prognostic RiskScore model we established has high prediction accuracy and survival prediction ability for brain metastases in NSCLC.