Developing a Statistical Modeling-Based Machine Learning Approach for Capturing Drug Dosing Using a Proton Pump Inhibitor Case.

Abstract

Objective: To develop a statistical model to capture medication dosing for proton pump inhibitors (PPIs) using structured data from electronic health records (EHR).

Methods: Medication data for PPIs was extracted from a single health care system EHR to develop a statistical model. Nearly 20 years' worth of PPI prescriptions were extracted and 25% of unique dosing regimens were manually labeled by two clinical pharmacists. Several machine learning models were trained and evaluated to predict dose. Training was applied to 70% of the unique dosing regimens. The remaining unique dosing regimens were tested and validated with standard regression metrics: root mean squared error (RMSE) and R-squared.

Results: A total of 17,271 distinct patients had orders for a PPI comprising 186,801 unique PPI orders. Distinct pairs built on medication descriptions and SIG combinations resulted in 10,739 unique entities. Clinical pharmacists manually labeled 2679 examples for medication entity extraction. Regression metrics (R-squared, RMSE) were chosen as metrics to evaluate model performance. A stacked ensembled model proved to have the best results with a 0.09 RMSE and an R-squared of 0.825.

Conclusion: The development of a statistical model to capture PPI dosing for both maintenance and complex dosing strategies was highly sensitive and accurate. A supervised learning prediction model helps overcome challenges in medication dosing identification by addressing concerns related to variability and complexity. Future strategies should focus on integrating unstructured data within the algorithm to further refine medication dosing capture.