Machine learning prediction models for mortality risk in sepsis-associated acute kidney injury: evaluating early versus late CRRT initiation

Abstract

Background: Sepsis-associated acute kidney injury (S-AKI) has a significant impact on patient survival, with continuous renal replacement therapy (CRRT) being a crucial intervention. However, the optimal timing for CRRT initiation remains controversial.

Methods: Using the MIMIC-IV database for model development and the eICU database for external validation, we analyzed patients with S-AKI to compare survival rates between early and late CRRT initiation groups. Propensity score matching was performed to address potential selection bias. Subgroup analyses stratified patients by disease severity using SOFA scores (low ≤10, medium 11-15, high >15) and creatinine levels (low ≤3 mg/dL, medium 3-5 mg/dL, high >5 mg/dL). Multiple machine learning models were developed and evaluated to predict patient prognosis, with Shapley Additive exPlanations (SHAP) analysis identifying key prognostic factors.

Results: After propensity score matching, late CRRT initiation was associated with improved survival probability, but led to increased hospital and ICU stays. Subgroup analyses showed consistent trends favoring late CRRT across all SOFA categories, with the most pronounced effect in high SOFA scores (>15, p = 0.058). The GBM model demonstrated robust predictive performance (average C-index 0.694 in validation and test sets). SHAP analysis identified maximum lactate levels, age, and minimum SpO2 as the strongest predictors of mortality, while CRRT timing showed relatively lower impact on outcome prediction.

Conclusion: While later initiation of CRRT in S-AKI patients was associated with improved survival, this benefit comes with increased healthcare resource utilization. The clinical parameters, rather than CRRT timing, are the primary determinants of patient outcomes, suggesting the need for a more personalized approach to CRRT initiation based on overall illness severity.

Keywords: CRRT timing; acute kidney injury; continuous renal replacement therapy; machine learning; mortality; sepsis.

Figure 1

Flowchart of the study cohort. Schematic representation of patient selection methodology from the MIMIC-IV 3.0 database, delineating the sequential filtration process and subsequent analytical stratification of sepsis-associated acute kidney injury cases into original and propensity-matched cohorts for comparative outcome analysis. CRRT, continuous renal replacement therapy; PSM, propensity score matching; KM, Kaplan–Meier; OS, overall survival.

Figure 2

Propensity score matching analysis and survival outcomes stratified by CRRT initiation timing. (A) Standardized mean differences of baseline covariates pre-and post-propensity score matching. (B) Propensity score distribution in early versus late CRRT cohorts before and after matching. (C–H) Kaplan–Meier survival estimates comparing early versus late CRRT initiation: MIMIC-IV cohort at (C) 14 days, (D) 28 days, and (E) 90 days; eICU cohort at (F) 14 days, (G) 28 days, and (H) 90 days post-CRRT initiation. Early versus late CRRT initiation was dichotomized at the median time interval from AKI onset to CRRT initiation. Shaded areas represent 95% confidence intervals. Log-rank test p values compare survival distributions between groups. At-risk tables display the number of patients under observation at specified time points.

Figure 3

Stratified analysis of 28-day survival outcomes by SOFA score and creatinine levels. Kaplan–Meier survival analyses comparing early versus late CRRT initiation stratified by disease severity markers. (A–C) SOFA score stratification: (A) Low SOFA (≤10), (B) Medium SOFA (11–15), and (C) High SOFA (>15). (D–F) First-day maximum serum creatinine stratification: (D) Low Cr (≤3 mg/dL), (E) Medium Cr (3–5 mg/dL), and (F) High Cr (>5 mg/dL). Late and early CRRT initiation groups are represented by blue and yellow curves, respectively. Shaded areas indicate 95% confidence intervals. Log-rank test p-values are shown for between-group comparisons. Numbers at risk are displayed below each curve at corresponding time points.

Figure 4

Machine learning model construction and performance analysis. (A) C-index heatmap comparing model performance across training, validation, and test datasets, with mean C-index values displayed. (B) UpSet plot illustrating the intersection of key model features with occurrence frequency greater than 3. (C) Time-dependent ROC curves for the GBM model at 14 days across training, validation, and test datasets. (D) Time-dependent ROC curves for the GBM model at 28 days across training, validation, and test datasets.

Figure 5

SHAP value analysis of features in the validation dataset. (A) Ranking of feature importance based on mean absolute SHAP values, showing top 10 clinical parameters and AKI-to-CRRT interval. (B) SHAP value distribution for key features with color gradient indicating feature values; points represent individual cases. (C) Feature value heatmap showing standardized (z-score) distribution across the validation cohort. (D–G) SHAP value interaction plots demonstrating the relationship between AKI-to-CRRT interval and: (D) maximum lactate levels, (E) age, (F) minimum SpO2, and (G) SOFA score. Color gradients represent AKI-to-CRRT interval levels (0–9).