Machine Learning Models for Point-of-Care Diagnostics of Acute Kidney Injury

Area of Science:

• Nephrology
• Artificial Intelligence
• Medical Informatics

Background:

• Early detection of acute kidney injury (AKI) often relies on baseline serum creatinine (SCr), a limitation for existing diagnostic algorithms.
• Point-of-care clinical features offer a potential alternative for AKI diagnosis when baseline SCr is unavailable.

Purpose of the Study:

• To develop and evaluate machine learning models for AKI diagnosis using readily available clinical features, independent of baseline SCr.
• To compare the diagnostic performance of these machine learning models against routine clinical diagnosis and a pre-existing computerized algorithm.

Main Methods:

• Retrospective cohort study using two datasets (n=2846 training, n=1331 testing) of patients with elevated SCr.
• Machine learning models were trained and tested using point-of-care features including laboratory data and physical readings.
• Model performance was assessed using Area Under the Receiver Operating Characteristic Curve (AUROC), precision, and F1 scores, with comparisons to clinician diagnosis and a benchmark algorithm.

Main Results:

• Machine learning models achieved AUROC values ranging from 0.67 to 0.74 on an independent test set.
• All machine learning models significantly outperformed routine clinician diagnosis (AUROC ~0.74 vs. 0.53, p < 0.05).
• A pre-existing algorithm requiring baseline SCr achieved a higher AUROC of 0.94, establishing a performance benchmark.

Conclusions:

• Machine learning models demonstrate superior accuracy in diagnosing AKI compared to routine clinical practice when baseline SCr is absent.
• Several advanced machine learning algorithms show comparable high performance in AKI diagnosis under these conditions.
• These findings highlight the potential of machine learning to improve AKI detection using accessible clinical data.