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

Background: Persistent sepsis-associated acute kidney injury (SA-AKI) shows poor clinical outcomes and remains a therapeutic challenge for clinicians. Early identification and prediction of persistent SA-AKI are crucial.

Objective: The aim of this study was to develop and validate an interpretable machine learning (ML) model that predicts persistent SA-AKI and to compare its diagnostic performance with that of C-C motif chemokine ligand 14 (CCL14) in a prospective cohort.

Methods: The study used 4 retrospective cohorts and 1 prospective cohort for model derivation and validation. The derivation cohort used the MIMIC-IV database, which was randomly split into 2 parts (80% for model construction and 20% for internal validation). External validation was conducted using subsets of the MIMIC-III dataset and e-ICU dataset, and retrospective cohorts from the intensive care unit (ICU) of Northern Jiangsu People's Hospital. Prospective data from the same ICU were used for validation and comparison with urinary CCL14 biomarker measurements. Acute kidney injury (AKI) was defined based on serum creatinine and urine output, using the Kidney Disease: Improving Global Outcomes (KDIGO) criteria. Routine clinical data within the first 24 hours of ICU admission were collected, and 8 ML algorithms were used to construct the prediction model. Multiple evaluation metrics, including area under the receiver operating characteristic curve (AUC), were used to compare predictive performance. Feature importance was ranked using Shapley Additive Explanations (SHAP), and the final model was explained accordingly. In addition, the model was developed into a web-based application using the Streamlit framework to facilitate its clinical application.

Results: A total of 46,097 patients with sepsis from multiple cohorts were enrolled for analysis. Among 17,928 patients with sepsis in the derivation cohort, 8081 patients (45.1%) showed progression to persistent SA-AKI. Among the 8 ML models, the gradient boosting machine (GBM) model demonstrated superior discriminative ability. Following feature importance ranking, a final interpretable GBM model comprising 12 features (AKI stage, ΔCreatinine, urine output, furosemide dose, BMI, Sequential Organ Failure Assessment score, kidney replacement therapy, mechanical ventilation, lactate, blood urea nitrogen, prothrombin time, and age) was established. The final model accurately predicted the occurrence of persistent SA-AKI in both internal (AUC=0.870) and external validation cohorts (MIMIC-III subset: AUC=0.891; e-ICU dataset: AUC=0.932; Northern Jiangsu People's Hospital retrospective cohort: AUC=0.983). In the prospective cohort, the GBM model outperformed urinary CCL14 in predicting persistent SA-AKI (GBM AUC=0.852 vs CCL14 AUC=0.821). The model has been transformed into an online clinical tool to facilitate its application in clinical settings.

Conclusions: The interpretable GBM model was shown to successfully and accurately predict the occurrence of persistent SA-AKI, demonstrating good predictive ability in both internal and external validation cohorts. Furthermore, the model was demonstrated to outperform the biomarker CCL14 in prospective cohort validation.