Incorporating temporal EHR data in predictive models for risk stratification of renal function deterioration.

Area of Science:

• Health Informatics
• Machine Learning
• Nephrology

Background:

• Electronic Health Records (EHRs) contain valuable temporal data for chronic disease management.
• Challenges in EHR data include irregular sampling and variable patient history lengths.
• Predicting kidney function decline is crucial for managing chronic kidney disease.

Purpose of the Study:

• To evaluate machine learning approaches for predictive modeling using temporal EHR data.
• To compare non-temporal aggregation with temporal modeling techniques.
• To assess the effectiveness of multi-task learning in capturing temporal dynamics for predicting kidney function loss.

Main Methods:

• Developed and compared three machine learning models: one non-temporal and two temporal.
• Temporal models differed in how they handled temporal information and missing data.
• Evaluated models using EHR data to predict loss of estimated glomerular filtration rate (eGFR).

Main Results:

• Incorporating temporal information significantly improved predictions of kidney function loss.
• The method of incorporating temporal dynamics is critical for model performance.
• Multi-task learning effectively captured time-varying predictor importance, outperforming other methods.

Conclusions:

• Temporal modeling in EHR data enhances predictions for chronic kidney disease progression.
• Multi-task learning provides a robust approach to leverage temporal EHR data for improved patient risk stratification.
• This approach can identify patients at high risk for short-term kidney function decline.