Machine learning-based prognostic model for local control prediction in lung cancer patients undergoing stereotactic body radiation therapy: a preliminary study.

PURPOSE: Predicting local recurrence after stereotactic body radiation therapy (SBRT) for lung cancer remains challenging. This study aims to develop a machine learning (ML)-based prognostic model for local control (LC) prediction by comparing different ML algorithms. METHODS: Clinical data from 158 lung cancer patients treated with SBRT were retrospectively analyzed. Six ML methods, including Boruta, RSF, GBM, LASSO, CoxBoost, and univariate Cox regression were systematically applied to perform variables selection, and the final prognostic model was constructed using multivariate Cox regression analysis. A benchmark model incorporating clinical stage plus treatment dose was also used for comparison. The performance of different models was evaluated using the C-index, Akaike Information Criterion (AIC), and time-dependent AUC. The optimal model was further evaluated comprehensively through receiver operating characteristic (ROC) analysis, calibration plots assessment, and decision curve analysis (DCA). The evaluation was validated through internal cross-validation to ensure robust reliability. RESULTS: The LASSO-Cox model achieved the highest C-index (0.718) and time-dependent AUC, shared the lowest AIC with the CoxBoost-Cox model, and outperformed the conventional stage plus treatment dose model (0.718 vs. 0.634). The LASSO-Cox model demonstrated moderate discriminatory ability, good calibration (predicted vs. observed outcomes), and clinical utility in DCA. CONCLUSIONS: ML demonstrates certain potential in LC prediction after SBRT. Nevertheless, it is worth noting that the study relies solely on internal validation and lacks external validation; therefore, further validation in independent cohorts is required before the model can be responsibly considered for clinical implementation.