Developing machine learning-driven acute kidney injury predictive models using non-standard EMRs in resource-limited...

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Nephrology

Background:

• Acute Kidney Injury (AKI) presents a global health challenge, particularly in resource-limited areas.
• Existing AKI prediction models often depend on serum creatinine (SCr) and standardized electronic medical records (EMRs).
• Non-standard EMRs and infrequent SCr testing in low-resource settings impede accurate AKI prediction.

Purpose of the Study:

• To develop and validate a machine learning model for AKI prediction using non-standard EMRs.
• To assess the model's efficacy in scenarios lacking SCr data.

Main Methods:

• A multicenter observational study involving 561,137 patients across 15 Chinese hospitals (2010-2016).
• Utilized the Light Gradient Boosting Machine (LightGBM) algorithm for predictive modeling.
• Evaluated model performance using metrics including AUC, precision, recall, specificity, and accuracy.

Main Results:

• 45,610 patients were diagnosed with AKI.
• The LightGBM model achieved high AKI prediction accuracy, with AUC values ranging from 0.860 to 0.986.
• Effective AKI prediction was demonstrated using non-standard EMRs, even without SCr data.

Conclusions:

• Non-standard EMRs are a valuable data source for AKI prediction, especially when SCr is unavailable.
• This approach is highly relevant for resource-limited settings, overcoming limitations of traditional biomarkers.
• Clinical features within non-standard EMRs can effectively compensate for the absence of SCr data in AKI prediction.