XGBoost-Cox model effectively predicts mortality risk in elderly male patients with acute kidney injury: a retrospective cohort study integrating nonlinear feature engineering and interpretable machine learning
- 1Department of Health Statistics， College of Military Preventive Medicine， Army Medical University （Third Military Medical University）， Chongqing
Abstract
Objective Elderly patients with acute kidney injury (AKI) face a high mortality risk that varies with renal function trajectories after AKI, whereas existing prediction models rely mainly on baseline variables and offer limited interpretability. Based on real-world follow-up data from elderly male patients with AKI, this study aimed to develop a mortality risk prediction framework that integrates nonlinear feature engineering with interpretable machine-learning survival models, and to evaluate its discrimination, calibration, clinical net benefit, and interpretability. Methods A single-center, retrospective cohort study was adopted to construct a prediction model based on older male inpatients who survived at least 7 days after AKI onset at Chinese PLA General Hospital between January 2008 and December 2025, with all-cause mortality as the endpoint outcome. Patients admitted between January 2008 and December 2018 constituted the model development cohort, and were randomly divided into a training set and an internal test set at a 7∶3 ratio; patients admitted between January 2019 and December 2025 constituted the temporal validation cohort. Variables collected included demographic characteristics, cardiovascular and hemodynamic status, renal function indicators, AKI severity and dynamic changes in renal function, hematological indicators, biochemical, nutritional and metabolic indices, electrolytes and minerals, medical history, comorbidities, and treatment-related factors. Restricted cubic splines (RCS) analysis was used to examine the nonlinear relationships between key continuous variables and mortality risk. Feature selection was performed by combining the Boruta algorithm, random survival forest (RSF), and clinical interpretability to construct 3 progressively expanded predictor sets. Six survival models were applied for mortality risk prediction, namely the Cox proportional hazards model, spline-extended Cox model, elastic-net Cox model, RSF, XGBoost-Cox, and DeepSurv. Discrimination was evaluated by the C-index and time-dependent AUC, and the calibration and clinical utility of the best-performing model were assessed by calibration curves and decision curve analysis. SHapley Additive exPlanations (SHAP) were further applied to interpret the best model globally and at the individual level. Results A total of 2 335 elderly male patients with AKI were finally included, of whom 1 912 died during follow-up and 423 survived. The training, internal test, and temporal validation sets comprised 775, 332, and 1 228 patients, respectively. BNP, BUN, and short-term post-AKI dynamic eGFR indicators showed nonlinear relationships with mortality risk (P<0.01). Age, nutritional status, cardiorenal functional load, hematological indicators, comorbidity burden, mechanical ventilation, and post-AKI changes in renal function were important predictors of mortality risk in elderly patients with AKI. The predictor set incorporating post-AKI dynamic renal function indicators revealed better overall predictive performance than models containing only baseline variables or AKI-onset variables. In the test set, the XGBoost-Cox model achieved the highest C-index and 3-year AUC, which were 0.749 (95%CI: 0.722 to 0.774) and 0.846 (95%CI: 0.802 to 0.891), respectively; RSF performed best in 30-day and 90-day predictions, with AUC values of 0.840 (95%CI: 0.785 to 0.888) and 0.829 (95%CI: 0.785 to 0.868), respectively. In temporal validation, the AUCs of the XGBoost-Cox model for 30-day, 90-day, 1-year, and 3-year mortality risk prediction were 0.797 (95%CI: 0.763 to 0.830), 0.781 (95%CI: 0.755 to 0.809), 0.802 (95%CI: 0.778 to 0.824), and 0.822 (95%CI: 0.788 to 0.853), respectively, indicating stable generalizability in the temporal validation cohort. The XGBoost-Cox model showed good discrimination, calibration generally consistent with observed risk at 3 years,and clinical net benefit, with renal functional load as its dominant predictor. Conclusion The prediction framework integrating nonlinear feature engineering with interpretable machine-learning survival models can effectively identify mortality risk in elderly male patients with AKI. The XGBoost-Cox model showed good discrimination in both internal testing and temporal validation, with generally consistent calibration and net benefit superior to treat-all and treat-none strategies for 3-year mortality risk prediction; SHAP analysis identified renal functional load as its dominant predictor, indicating that the model's predictions are interpretable.
Related
- Integrating 72-h BUN/Cr trajectories with interpretable machine learning for 28-day mortality after acute stroke: a dual-center landmark study.
- Development and interpretable validation of a machine-learning model for edentulism risk in US adults : A cross-sectional study integrating five algorithms and dietary data from NHANES 2009–2014
- Development and validation of an interpretable machine learning model for early prediction of mortality in ICU patients with chronic obstructive pulmonary disease and heart failure
- An interpretable machine learning model integrating mucosal and cellular immunity to predict SARS-CoV-2 reinfection: a multicenter cohort study.
- Modelled mortality benefits of reducing fibrinolysis resistance in sepsis: integrating observational and ex vivo data
BibTeX
@article{qiuyue2026xgboostcox,
title = {{XGBoost-Cox model effectively predicts mortality risk in elderly male patients with acute kidney injury: a retrospective cohort study integrating nonlinear feature engineering and interpretable machine learning}},
author = {Song Qiuyue and Wu Yazhou},
journal = {陆军军医大学学报},
year = {2026},
month = oct,
doi = {10.16016/j.2097-0927.202607040},
url = {https://doi.org/10.16016/j.2097-0927.202607040},
note = {CC BY},
}
RIS
TY - JOUR
TI - XGBoost-Cox model effectively predicts mortality risk in elderly male patients with acute kidney injury: a retrospective cohort study integrating nonlinear feature engineering and interpretable machine learning
AU - Song Qiuyue
AU - Wu Yazhou
PY - 2026
DA - 2026/10/30
JO - 陆军军医大学学报
DO - 10.16016/j.2097-0927.202607040
UR - https://doi.org/10.16016/j.2097-0927.202607040
AB - Objective Elderly patients with acute kidney injury (AKI) face a high mortality risk that varies with renal function trajectories after AKI, whereas existing prediction models rely mainly on baseline variables and offer limited interpretability. Based on real-world follow-up data from elderly male patients with AKI, this study aimed to develop a mortality risk prediction framework that integrates nonlinear feature engineering with interpretable machine-learning survival models, and to evaluate its discrimination, calibration, clinical net benefit, and interpretability. Methods A single-center, retrospective cohort study was adopted to construct a prediction model based on older male inpatients who survived at least 7 days after AKI onset at Chinese PLA General Hospital between January 2008 and December 2025, with all-cause mortality as the endpoint outcome. Patients admitted between January 2008 and December 2018 constituted the model development cohort, and were randomly divided into a training set and an internal test set at a 7∶3 ratio; patients admitted between January 2019 and December 2025 constituted the temporal validation cohort. Variables collected included demographic characteristics, cardiovascular and hemodynamic status, renal function indicators, AKI severity and dynamic changes in renal function, hematological indicators, biochemical, nutritional and metabolic indices, electrolytes and minerals, medical history, comorbidities, and treatment-related factors. Restricted cubic splines (RCS) analysis was used to examine the nonlinear relationships between key continuous variables and mortality risk. Feature selection was performed by combining the Boruta algorithm, random survival forest (RSF), and clinical interpretability to construct 3 progressively expanded predictor sets. Six survival models were applied for mortality risk prediction, namely the Cox proportional hazards model, spline-extended Cox model, elastic-net Cox model, RSF, XGBoost-Cox, and DeepSurv. Discrimination was evaluated by the C-index and time-dependent AUC, and the calibration and clinical utility of the best-performing model were assessed by calibration curves and decision curve analysis. SHapley Additive exPlanations (SHAP) were further applied to interpret the best model globally and at the individual level. Results A total of 2 335 elderly male patients with AKI were finally included, of whom 1 912 died during follow-up and 423 survived. The training, internal test, and temporal validation sets comprised 775, 332, and 1 228 patients, respectively. BNP, BUN, and short-term post-AKI dynamic eGFR indicators showed nonlinear relationships with mortality risk (P<0.01). Age, nutritional status, cardiorenal functional load, hematological indicators, comorbidity burden, mechanical ventilation, and post-AKI changes in renal function were important predictors of mortality risk in elderly patients with AKI. The predictor set incorporating post-AKI dynamic renal function indicators revealed better overall predictive performance than models containing only baseline variables or AKI-onset variables. In the test set, the XGBoost-Cox model achieved the highest C-index and 3-year AUC, which were 0.749 (95%CI: 0.722 to 0.774) and 0.846 (95%CI: 0.802 to 0.891), respectively; RSF performed best in 30-day and 90-day predictions, with AUC values of 0.840 (95%CI: 0.785 to 0.888) and 0.829 (95%CI: 0.785 to 0.868), respectively. In temporal validation, the AUCs of the XGBoost-Cox model for 30-day, 90-day, 1-year, and 3-year mortality risk prediction were 0.797 (95%CI: 0.763 to 0.830), 0.781 (95%CI: 0.755 to 0.809), 0.802 (95%CI: 0.778 to 0.824), and 0.822 (95%CI: 0.788 to 0.853), respectively, indicating stable generalizability in the temporal validation cohort. The XGBoost-Cox model showed good discrimination, calibration generally consistent with observed risk at 3 years,and clinical net benefit, with renal functional load as its dominant predictor. Conclusion The prediction framework integrating nonlinear feature engineering with interpretable machine-learning survival models can effectively identify mortality risk in elderly male patients with AKI. The XGBoost-Cox model showed good discrimination in both internal testing and temporal validation, with generally consistent calibration and net benefit superior to treat-all and treat-none strategies for 3-year mortality risk prediction; SHAP analysis identified renal functional load as its dominant predictor, indicating that the model's predictions are interpretable.
SN - 2097-0927
ER -