Machine Learning-Based Mortality Risk Prediction Models in Patients with Sepsis-Associated Acute Kidney Injury: A Systematic Review

Xu Li^1,2Xu Hu¹Huiting Xu^2,3Pin Yu^1*Hailing Ju^4*

• ¹First Affiliated Hospital of Soochow University, Suzhou, China
• ²Tongji University, Shanghai, China
• ³Qingpu Branch of Zhongshan Hospital Affiliated to Fudan University, Shanghai, China
• ⁴Shanghai Tenth People's Hospital, Shanghai, China

Background: Machine learning (ML) models are increasingly utilized to predict mortality in patients with sepsis-associated acute kidney injury (SA-AKI), frequently surpassing traditional scoring systems. Despite their efficacy, inconsistencies in model quality remain a concern. This review aims to evaluate existing ML-based SA-AKI mortality prediction models, with a focus on development quality, methodological rigor, and predictive performance. Objective: To systematically assess ML-based mortality risk prediction models for SA-AKI patients. Methods: A comprehensive literature search on ML-based SA-AKI mortality prediction models was conducted across PubMed, Cochrane, Embase, and Web of Science from the inception of these databases until July 2025. Two researchers independently screened the literature, extracted data, and assessed model quality employing the Prediction Model Risk of Bias Assessment Tool for Artificial Intelligence. Results: Nine studies were included, all of which entailed model development and validation phases; five were solely internally validated while four underwent external validation as well. The studies utilized 18 different algorithms, with Random Forest and Extreme Gradient Boosting being the most prevalent. The majority of the studies employed K-nearest neighbor or Multiple Imputation by Chained Equations for handling missing values and utilized Recursive Feature Elimination, Least Absolute Shrinkage and Selection Operator, and Boruta's algorithm for feature selection. Seven studies assessed model calibration performance. The Area Under the Curve (AUC) for the training sets generally ranged from 0.75 to 0.99, which decreased to 0.70 to 0.87 during internal validation. Extreme Gradient Boosting consistently showed robust performance in external validation. The final predictors encompassed six principal categories: demographic information, vital signs, laboratory tests, disease severity, comorbidities, and interventions. Conclusions: ML models demonstrate promising performance and applicability in predicting mortality risk in SA-AKI patients, with consistent core predictors. Nevertheless, most studies exhibit a potential risk of bias. Future efforts should aim to enhance the standardization of data processing, feature selection, and validation processes. Additionally, there is a need to focus on the construction of prospective models based on early variables, and to ensure the interpretability and clinical integration of the models to facilitate their practical application in healthcare workflows.