Optimal sampling for positive only electronic health record data.

Area of Science:

• Biomedical Informatics
• Machine Learning in Healthcare
• Health Data Science

Background:

• Accurate patient phenotyping from electronic health records (EHR) is vital for research.
• Phenotype assessment is resource-intensive, necessitating efficient training data selection.
• Existing methods may not optimize for limited sample sizes or model misspecification.

Purpose of the Study:

• To develop a procedure for selecting optimal EHR subsamples for training classification models.
• To minimize mean-squared error (MSE) in phenotyping or classification tasks.
• To incorporate "positive only" data for improved training efficiency.

Main Methods:

• A novel sampling procedure designed to tailor the best training subsample from EHR data.
• The method minimizes mean-squared phenotyping/classification error (MSE).
• Incorporation of "positive only" information, an approximation of true disease status.

Main Results:

• Theoretical justification for the optimality of the proposed sampling procedure in terms of MSE.
• Demonstrated performance gains through simulations and a real-data example.
• The method proves effective even when the classification model is misspecified.

Conclusions:

• The proposed method offers an efficient approach to EHR data subsampling for model training.
• It effectively reduces phenotyping/classification error, especially with limited data.
• The procedure is robust and performs well under various conditions, including model misspecification.