Development and Validation of Machine Learning Models to Identify Emergency Department Patients at Increased Risk of...

Area of Science:

• Nephrology
• Data Science
• Clinical Informatics

Background:

• Acute kidney injury (AKI) is a significant clinical challenge with high mortality and morbidity.
• Existing AKI prediction models often lack generalizability to emergency department (ED) settings, especially for discharged patients.
• Accurate early prediction of AKI is crucial for timely intervention and mitigating adverse outcomes.

Purpose of the Study:

• To develop and validate machine learning (ML) models for predicting new or progressive AKI within 72 hours of ED departure.
• To address the challenge of missing outcome data for patients discharged from the ED.
• To enhance the real-world applicability of AKI prediction models.

Main Methods:

• Retrospective, multicenter study of adult ED visits using routinely collected demographic, clinical, and laboratory data.
• Extreme gradient boosting algorithms employed with four methods to handle missing outcome data.
• Model performance evaluated using cross-validation and external temporal validation with AUC, precision, recall, and calibration.

Main Results:

• Models demonstrated robust predictive performance for any AKI (AUC 0.81-0.82) and severe AKI (AUC 0.87-0.88).
• Inverse probability weighting proved effective for handling missing data, providing accurate risk estimates for hospitalized and discharged patients.
• Consistent performance observed across diverse subgroups and ED sites.

Conclusions:

• Machine learning models trained on ED data offer reliable early AKI prediction.
• These models support clinical decision-making for a broad patient spectrum, including those discharged.
• The study advances ML model usability in real-world settings by including discharged patients and estimating ongoing kidney risk.