A systematic review of artificial intelligence algorithms for predicting acute kidney injury.

Abstract

Acute kidney injury (AKI) increases mortality and costs in hospitalized patients. New methods for early AKI identification have been developed with targeted biomarkers and electronic health records data analysis. Machine learning (ML) use in diagnostics and health data analysis has recently increased. We performed a systematic review to analyze the use of ML for AKI prediction in hospitalized adults. Tubmed, EMBASE, Cochrane, and Web of Science databases were searched until 31st March of 2023. English-language studies using ML in adults for AKI prediction were included using predetermined eligibility search terms such as acute kidney injury, machine learning, and artificial intelligence. Two reviewers evaluated the publications' titles, abstracts, and full texts separately and obtained appropriate data. The main outcome was an area under the curve (AUC) result of at least 0.70. Ten studies in 102 articles were included involving 242,251 patients. Deep learning (AUC 0.907 in critical care AKI; AUC 0.797 in hospitalized patients AKI) was similar to Logistic regression (AUC 0.877 in critical care AKI; AUC 0.789 in hospitalized patients). Decision tree constructions had similar AUC. In this review, most ML models analyzed fulfilled the main outcome. AKI is multifactorial; however, ML performed well with different etiologies, such as cardiac-related AKI, drug-related AKI, and critical care patients. Overfitting data and constructing black box models are limitations that might jeopardize the generalization and comprehension of the results. Most studies were single-center, and three manuscripts used the same database with a predominantly Caucasian population, resulting in a lack of diversity and reducing external generalization. In conclusion, ML could effectively predict AKI in hospitalized adults. Future directions rely on including a more diverse population and completing prospective and controlled trials.