An explainable attention-gated stacking ensemble model for predicting survival in non-small cell lung cancer: development and validation.

BACKGROUND

Non-small cell lung cancer (NSCLC) remains the leading cause of cancer-related mortality worldwide, and accurate survival prediction is essential for individualized treatment planning. Existing prognostic models often rely on single algorithms with limited capacity to capture the complexity inherent in clinical data. This study aimed to develop, validate, and interpret an attention-gated stacking ensemble model for predicting 1-, 3-, 5-year overall survival (OS) in NSCLC patients.

METHODS

A retrospective cohort of 67,400 NSCLC patients from the Surveillance, Epidemiology, and End Results (SEER) database (2004-2021) was analyzed. A dual-filtering strategy combining Pearson and Spearman correlation analysis identified 11 prognostic features. An attention-gated stacked ensemble model was constructed, consisting of eleven diverse base learners plus a feature attention-gated mechanism, which ultimately generated survival probability estimate of 1, 3, and 5 years. Model performance was assessed using area under the receiver operating characteristic curve (AUC), confusion matrices, calibration curves with Brier scores, and decision curve analysis (DCA). An independent external validation cohort of 623 patients from Xinqiao Hospital (2015-2023) was used to evaluate generalizability. SHapley Additive exPlanations (SHAP) analysis was employed for model explanation. A web-based tool was developed for clinical application.

RESULTS

On the internal test set, the attention-gated stacking ensemble model achieved AUC values of 0.799, 0.799, and 0.778 for 1-, 3-, and 5-year OS prediction, respectively, outperforming all eleven individual base learners and a conventional stacking model without the attention gate. On external validation, AUC values reached 0.817, 0.807, and 0.825, with Brier scores of 0.127, 0.107, and 0.095, demonstrating robust cross-population generalizability. SHAP analysis identified metastasis (M) stage, tumor size, and lymph node (N) stage as the three most influential predictors, and the learned attention gate weights exhibited concordant importance rankings. A publicly accessible web-based calculator was deployed to facilitate personalized survival estimation at the point of care.

CONCLUSIONS

The proposed attention-gate stacking ensemble model provides accurate and explainable survival predictions for NSCLC patients, outperforming conventional single-algorithm and standard stacking approaches. The integration of a feature-wise attention gate mechanism enhances both predictive performance and clinical transparency, supporting its potential utility in personalized oncological decision-making. The deployed web-based clinical decision support tool further bridges the gap between algorithmic development and bedside application.