JMIR Publications

ABSTRACT

Background:

Sepsis-associated acute kidney injury (SA-AKI) is a frequent complication in ICU patients and significantly increases both mortality rates and the risk of chronic kidney disease. While most research has focused on broad risk prediction, there is a notable lack of detailed stratification on the basis of the severity of kidney injury, limiting its clinical applicability.

Objective:

This study aimed to identify critical time points in SA-AKI progression through dynamic trend analysis and to develop and validate machine learning-based risk stratification models for moderate-to-severe SA-AKI.

Methods:

We utilized data from the MIMIC-IV 3.0 (n=12,842) and eICU-CRD 2.0 (n=15,767) databases for model development and external validation, respectively. AKI stages were analyzed at 48 hours and 3, 7, and 28 days post-ICU admission, with 48 hours (acute phase) and 7 days (subacute phase) considered pivotal time points for model construction. Clinical data from the first 24 hours of ICU admission were processed using LightGBM, with SHAP values applied for feature selection and model interpretability. Six machine learning algorithms were compared, and the best-performing model was selected on the basis of the AUC, calibration curves, and decision curve analysis (DCA). Internal validation was conducted via fivefold cross-validation, whereas external validation was used to assess model generalizability. SHAP values and partial dependence plots (PDPs) were used to elucidate the effects of key features on the predictions.

Results:

Dynamic trend analysis revealed a 46.7% increase in severe SA-AKI cases (stages 2–3) during the acute phase (1–2 days) (P<.01), with a 523.5% increase in Stage 3 cases (P<.01). In the subacute phase (2–7 days), the proportion of severe cases increased by 24.7% (P=.02), whereas in the stable phase (7–28 days), only a marginal 3.9% increase was detected (P=.08). Among the six machine learning models, LightGBM demonstrated superior performance, achieving AUCs of 0.862 in the training cohort and 0.770 in the external validation set for the 48-hour model and AUCs of 0.922 and 0.720, respectively, for the 7-day model. SHAP analysis identified urine output, mechanical ventilation use, the SOFA score, and nephrotoxic drug use as key predictors across both models. The PDPs revealed optimal thresholds for modifiable risk factors. DCA confirmed that LightGBM provided consistent clinical benefits across a range of thresholds.

Conclusions:

This study is the first to develop stratified prediction models for moderate-to-severe SA-AKI based on dynamic disease progression. The models effectively identify high-risk patients early and support phase-specific intervention strategies, providing a robust foundation for the clinical management of SA-AKI in critically ill patients.