A Web-Based Clinical Decision Support System for Cardio-Kidney Risk Stratification in Critically Ill Patients with Diabetic Kidney Disease: Development, SHAP Interpretability, and Multi-Center External Validation

<title>Abstract</title> <p> <bold>Background:</bold> Diabetic kidney disease (DKD) is a major cardiovascular risk factor associated with progressive kidney failure and adverse cardiovascular events. Critically ill patients with DKD have poor prognoses; however, data-driven clinical decision support tools integrating renal and cardiovascular risks are lacking. <bold>Methods:</bold> Using the Medical Information Mart for Intensive Care IV database, we developed a web-based clinical decision support system (CDSS) for cardio-kidney risk stratification. The model used LASSO-Cox regression on 4,478 ICU patients with DKD (7:3 training/validation split). SHAP analysis enhanced model interpretability, addressing a key barrier to clinical application. External validation used in-hospital mortality data from the eICU-CRD database; after applying the inclusion/exclusion criteria (ICU stays &lt;24 h), 2,286 patients remained (277 deaths). Because the eICU-CRD lacked post-discharge follow-up, we built a logistic model for in-hospital mortality using the same 14 predictors for direct endpoint comparison. Multiple imputation (MICE) generated m=20 datasets to handle the missing data. <bold>Results:</bold> The CDSS incorporated 14 predictors: age, albumin, hemoglobin, BUN, INR, potassium, magnesium, anion gap, heart failure, severe liver disease, RRT, and ICU admission, achieving an internal validation AUC of 0.816-0.844. SHAP analysis identified albumin (0.42), age (0.41), and heart failure (0.25) as the top contributors, validating the model's cardio-renal risk capture. In the sensitivity analysis using the MIMIC-IV hospital-mortality logistic model, the development cohort achieved an AUC of 0.739 (95% CI: 0.680-0.797), calibration slope of 0.709, and HL p=0.0007. External validation in the eICU-CRD cohort yielded an AUC of 0.784 (95% CI: 0.732-0.837), Brier score of 0.085 (95% CI: 0.071-0.104), calibration intercept of 0.476 (95% CI: -0.207-1.199), calibration slope of 1.366 (95% CI: 1.023-1.753), and Hosmer-Lemeshow p=0.082. Decision curve analysis confirmed a positive net clinical benefit across the 0.02-0.50 thresholds. <bold>Conclusion:</bold> This externally validated interpretable CDSS supports real-time bedside cardio-kidney risk stratification. Local intercept recalibration is recommended prior to implementation. </p>