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

Area of Science:

• Pharmacometrics
• Health Informatics
• Machine Learning in Healthcare

Background:

• Proton pump inhibitors (PPIs) are widely prescribed, but accurately capturing their dosing from electronic health records (EHR) presents challenges due to variability and complexity.
• Structured EHR data offers potential for developing automated medication dosing models.

Purpose of the Study:

• To develop and evaluate a statistical model for capturing proton pump inhibitor (PPI) medication dosing using structured data from electronic health records (EHR).

Main Methods:

• Extracted nearly 20 years of PPI prescription data from a single healthcare system's EHR.
• Manually labeled 25% of unique dosing regimens by clinical pharmacists for model training and validation.
• Trained and evaluated several machine learning models, including a stacked ensemble model, using regression metrics (RMSE, R-squared).

Main Results:

• The study analyzed 17,271 patients and 186,801 unique PPI orders, identifying 10,739 unique medication entities.
• A stacked ensemble model achieved the best performance with a Root Mean Squared Error (RMSE) of 0.09 and an R-squared value of 0.825.
• The model demonstrated high sensitivity and accuracy in capturing PPI dosing.

Conclusions:

• Developed a highly sensitive and accurate statistical model for capturing PPI dosing, including complex strategies.
• Supervised learning models can effectively address challenges in medication dosing identification.
• Future work should integrate unstructured EHR data to further enhance medication dosing capture accuracy.