Related Experiment Video
Updated: Jan 8, 2026
Constructing and Visualizing Models using Mime-based Machine-learning Framework
Published on: July 22, 2025
Literature-informed ensemble machine learning for three-year diabetic kidney disease risk prediction in type 2
Ayla M Tourkmani1, Turki J Al-Harbi1, Ahmad Abdullah Alghamdi2
1Department of Family and Community Medicine, Prince Sultan Military Medical City, Riyadh, Saudi Arabia.
Introduction:
Diabetic kidney disease (DKD) and diabetic nephropathy (DN) affect around 40% of diabetic patients but lack accurate risk prediction tools that include social determinants and demographic complexity. We developed and validated an ensemble machine learning model for three-year DKD/DN risk prediction with deployment readiness.
Methods:
We analysed 18 742 eligible adult type 2 diabetic patients from Prince Sultan Military Medical City (PSMMC) registry between 2019 and 2024 in Riyadh, Saudi Arabia. Using temporal patient-level splitting, we developed a stacked ensemble model (LightGBM + CoxBoost) with several features including multiple literature-informed imputed variables including family history, non-steroidal anti-inflammatory drug (NSAID) use, socioeconomic deprivation, diabetic retinopathy severity, and antihypertensive medications, imputed via Bayesian multiple imputation by chained equations (MICE) with external study priors. Primary outcome was incident/progressive DKD/DN within 3 years' timeframe. We assessed discrimination, calibration, model utilisation, and algorithmic fairness.
Results:
The final model achieved excellent discrimination (receiver operating characteristic [AUROC] of 0.852, 95% CI 0.847-0.857) and near-perfect calibration (slope 0.98, intercept -0.012) on multi-trial validation. Decision curve evaluation demonstrated superior net benefit (+22 events prevented per 1000 patients at 10% threshold) compared to treat-all strategies. Bootstrap validation showed minimal optimism in discrimination (C-statistic optimism = 0.005). No algorithmic bias was detected across demographic subgroups (maximum |Δ-AUROC| = 0.010). Prior sensitivity analysis confirmed validity and significance (AUROC variation ≤0.008). The model was engineered and deployed as an interactive web-based application (https://nephrarisk.streamlit.app/).
Conclusions:
Our developed and demonstrated model provided accurate and well-fair DKD/DN risk prediction with excellent calibration, allowing for better decision making with deployment as a web-based research tool and framework for future prospective clinical validation. Further validation and testing are warranted from different centres and healthcare systems to increase confidence and dissemination of our model findings for better utilisation purposes in the future.
