Development and external validation of machine learning models for predicting mortality in patients with lung cancer: a multicenter retrospective cohort study
- 1Hubei University of Chinese Medicine
- 2Hubei Provincial Hospital of Traditional Chinese Medicine
- 3School of Public Health, Shanghai Jiao Tong University
Abstract
Lung cancer remains the leading cause of cancer mortality worldwide. Accurate prognostic prediction can support clinical decision-making and resource allocation, yet many existing models use limited predictors and lack independent validation. We developed and externally validated machine-learning models to predict mortality in patients with lung cancer using routinely collected clinical, laboratory, and treatment-related variables. We conducted a multicenter retrospective cohort study including 2,783 patients with primary lung cancer from seven hospitals in China (January 2011 to end of data collection). Six centers formed the development cohort (n = 2,398; 70% training [n = 1,678] and 30% internal test [n = 720]), and one independent center served as external validation (n = 385). Predictors were selected using Boruta followed by LASSO. We compared six algorithms (logistic regression, random forest, support vector machine, gradient boosting machine, XGBoost, and neural network). Discrimination, calibration, and clinical utility were assessed using AUC, calibration plots/Brier score, and decision curve analysis. Model interpretation used SHAP. Seventeen predictors were retained from 51 candidates, spanning treatment (chemotherapy, intervention), tumor stage (T stage, M stage, stage group, distant metastasis), inflammatory markers (white blood cell count, neutrophil percentage, monocyte-to-lymphocyte ratio), nutrition/hematology (albumin, red blood cell count), tumor markers (CYFRA21-1, CEA), coagulation (PT, APTT), and electrolytes (sodium, potassium). Random forest performed best in the internal test set (AUC 0.861, 95% CI 0.835–0.886). In external validation, performance attenuated but remained acceptable (AUC 0.749, 95% CI 0.693–0.799; accuracy 72.7%; sensitivity 80.0%; specificity 49.4%), with reasonable calibration (Brier score 0.159) and net benefit across relevant thresholds. SHAP identified T stage, chemotherapy status, white blood cell count, albumin, and M stage as key contributors. We developed a web-based risk calculator. A multicenter machine-learning model integrating routinely available variables achieved moderate externally validated performance for mortality prediction in lung cancer. The accompanying web-based calculator may facilitate individualized risk stratification, pending further validation and potential recalibration in new settings.
Related
- Development and validation of a nomogram for predicting chest drainage after CT-guided percutaneous lung biopsy: a multicenter retrospective study
- Development and Validation of Daping Orthopedics Operative Risk Scoring System for Geriatric Hip Fracture (DOORS‐GHF) for Predicting Postoperative Outcomes in Geriatric Hip Fracture Patients: A Multicenter Retrospective Comparative Study
- Development and internal validation of a Septic Shock Clinical Score System (SCSS) for predicting 28-day mortality in pediatric patients: a single-center retrospective study.
- Development and validation of a novel metabolic-inflammatory health-related nomogram for predicting in-hospital mortality in older patients with type 2 diabetes mellitus combined with cardiovascular disease
- Development and internal validation of a nomogram for predicting flexible bronchoscopy utilization in children with lobar pneumonia caused by Mycoplasma pneumoniae: a single-center retrospective cohort study
BibTeX
@article{liu2026development,
title = {{Development and external validation of machine learning models for predicting mortality in patients with lung cancer: a multicenter retrospective cohort study}},
author = {Lihui Liu and Pan Zuo and Chunlan Yu and Xinhai Shen and Baoping Luo and Siqi Zhang and Ting Cao and Yong Zhou and Kaiwen Hu and Yong Cai},
journal = {Scientific Reports},
year = {2026},
month = jul,
doi = {10.1038/s41598-026-47227-w},
url = {https://doi.org/10.1038/s41598-026-47227-w},
note = {CC BY},
}
RIS
TY - JOUR
TI - Development and external validation of machine learning models for predicting mortality in patients with lung cancer: a multicenter retrospective cohort study
AU - Lihui Liu
AU - Pan Zuo
AU - Chunlan Yu
AU - Xinhai Shen
AU - Baoping Luo
AU - Siqi Zhang
AU - Ting Cao
AU - Yong Zhou
AU - Kaiwen Hu
AU - Yong Cai
PY - 2026
DA - 2026/07/01
JO - Scientific Reports
DO - 10.1038/s41598-026-47227-w
UR - https://doi.org/10.1038/s41598-026-47227-w
AB - Lung cancer remains the leading cause of cancer mortality worldwide. Accurate prognostic prediction can support clinical decision-making and resource allocation, yet many existing models use limited predictors and lack independent validation. We developed and externally validated machine-learning models to predict mortality in patients with lung cancer using routinely collected clinical, laboratory, and treatment-related variables. We conducted a multicenter retrospective cohort study including 2,783 patients with primary lung cancer from seven hospitals in China (January 2011 to end of data collection). Six centers formed the development cohort (n = 2,398; 70% training [n = 1,678] and 30% internal test [n = 720]), and one independent center served as external validation (n = 385). Predictors were selected using Boruta followed by LASSO. We compared six algorithms (logistic regression, random forest, support vector machine, gradient boosting machine, XGBoost, and neural network). Discrimination, calibration, and clinical utility were assessed using AUC, calibration plots/Brier score, and decision curve analysis. Model interpretation used SHAP. Seventeen predictors were retained from 51 candidates, spanning treatment (chemotherapy, intervention), tumor stage (T stage, M stage, stage group, distant metastasis), inflammatory markers (white blood cell count, neutrophil percentage, monocyte-to-lymphocyte ratio), nutrition/hematology (albumin, red blood cell count), tumor markers (CYFRA21-1, CEA), coagulation (PT, APTT), and electrolytes (sodium, potassium). Random forest performed best in the internal test set (AUC 0.861, 95% CI 0.835–0.886). In external validation, performance attenuated but remained acceptable (AUC 0.749, 95% CI 0.693–0.799; accuracy 72.7%; sensitivity 80.0%; specificity 49.4%), with reasonable calibration (Brier score 0.159) and net benefit across relevant thresholds. SHAP identified T stage, chemotherapy status, white blood cell count, albumin, and M stage as key contributors. We developed a web-based risk calculator. A multicenter machine-learning model integrating routinely available variables achieved moderate externally validated performance for mortality prediction in lung cancer. The accompanying web-based calculator may facilitate individualized risk stratification, pending further validation and potential recalibration in new settings.
SN - 2045-2322
ER -