Objective To explore the predictive value of machine learning model based on treatment regimens, CT imaging features and clinical data for the short-term efficacy of advanced non-small cell lung cancer (NSCLC).

Methods A retrospective collection was made of 151 patients with advanced NSCLC from January 2020 to April 2025. Among them, 96 patients from January 2020 to December 2023 were randomly divided into the training set andtest set at a ratio of 6:4. Fifty-five patients from January 2024 to April 2025 were used as the independent test set. The treatment regimens, clinical baselines and CT imaging features were collected. The random forest, extreme gradient boosting tree, lightweight gradient boosting machine (LGBM) and logistic regression models were constructed. ROC curve, calibration curve and decision curve analysis (DCA) were used to evaluate the performance of the models.

Results  The AUC of the LGBM model in the training set, test set 1 and test set 2 was 0.741, 0.750 and 0.727, respectively, and the AUC fluctuation range was 0.023, which was smaller than that of Random forest (0.235) and XGBoost (0.312). The accuracies of test Set 1 and test Set 2 were 0.759 and 0.618, respectively, and the sensitivities were 0.882 and 0.828, respectively. The DeLong test showed that there was no statistically significant difference in the AUC between LGBM and logistic regression in the three datasets (P ＞ 0.05). Combined with the calibration curve and DCA, it suggested that its stability and clinical net benefit were better.

Conclusions The machine learning model integrating PD-1 treatment status, CT image features and clinical baseline can be used for the short-term efficacy prediction of advanced NSCLC. Among them, the LGBM model shows relatively stable performance in internal and external tests. The results of SHAP analysis suggests that smoking history, long diameter of the lesion and PD-1 treatment status are the important factors affecting the judgment of therapeutic effect, which can provide a reference for individualized treatment decisions.