JMIR Publications

ABSTRACT

Background:

Most artificial intelligence-based research on acute kidney injury prediction has focused on intensive care unit settings, limiting their generalizability to general wards. The lack of standardized acute kidney injury definitions and reliance on intensive care unit further hinder the clinical applicability of these models.

Objective:

To develop and validate a machine learning-based framework for early prediction of acute kidney injury and acute kidney disease, applicable to general ward patients, using a refined operational definition of acute kidney injury.

Methods:

This retrospective cohort study utilized data from the three different hospitals. Acute kidney injury and acute kidney disease were defined based on the Kidney Disease: Improving Global Outcomes criteria with refinements to the baseline serum creatinine determination and the minimum serum creatinine increase required for the relative criteria. The primary outcome was the development of models for early acute kidney injury and acute kidney disease prediction.

Results:

The final analysis included 135,068 patients (mean [standard deviation] age, 61.9 [16.3] years; 43% female). The incidence of acute kidney injury was 8.0% and 7.3% in the internal and external cohorts, respectively. The acute kidney injury and acute kidney disease prediction models achieved the area under the receiver operating characteristic curve of 0.913 (95% confidence interval, 0.909-0.918) and 0.781 (95% confidence interval, 0.765-0.796), respectively.

Conclusions:

The refined acute kidney injury definition significantly reduced the classification of patients with transient serum creatinine fluctuations as acute kidney injury cases compared to the previous criteria.