Explainable Machine Learning Model for Predicting Persistent Sepsis-Associated Acute Kidney Injury: Development and...

Area of Science:

• Nephrology
• Critical Care Medicine
• Data Science in Healthcare

Background:

• Persistent sepsis-associated acute kidney injury (SA-AKI) presents significant clinical challenges and poor outcomes.
• Early and accurate prediction of persistent SA-AKI is critical for timely intervention.

Purpose of the Study:

• To develop and validate an interpretable machine learning (ML) model for predicting persistent SA-AKI.
• To compare the diagnostic performance of the ML model against the urinary biomarker C-C motif chemokine ligand 14 (CCL14).

Main Methods:

• Utilized multiple retrospective and prospective cohorts, including MIMIC-IV, MIMIC-III, and e-ICU databases.
• Developed and validated 8 ML algorithms, selecting a gradient boosting machine (GBM) model based on performance metrics.
• Employed Shapley Additive Explanations (SHAP) for model interpretability and developed a web-based application for clinical use.

Main Results:

• The final interpretable GBM model, using 12 key clinical features, demonstrated high accuracy in predicting persistent SA-AKI across internal and external validation cohorts (AUCs ranging from 0.870 to 0.983).
• In a prospective cohort, the GBM model showed superior predictive performance compared to urinary CCL14 (AUC=0.852 vs. 0.821).

Conclusions:

• An interpretable GBM model effectively predicts persistent SA-AKI with strong validation across diverse cohorts.
• The developed ML model offers a promising alternative and outperforms the CCL14 biomarker for predicting persistent SA-AKI in clinical settings.