Predicting the physiological effects of multiple drugs using electronic health record

Area of Science:

• Pharmacology
• Biomedical Informatics
• Machine Learning

Background:

• Existing computational models for drug-drug interactions (DDIs) primarily focus on pairwise interactions.
• These models often neglect crucial patient-specific information, limiting their clinical applicability.
• Predicting complex physiological effects of multiple drug interactions remains a significant challenge.

Purpose of the Study:

• To develop and validate machine learning models for predicting physiological effects of polypharmacy (two or more drugs).
• To incorporate patient data from electronic health records (EHRs) into DDI prediction models.
• To identify key features influencing drug interaction outcomes.

Main Methods:

• Utilized the MIMIC-IV database, a large, publicly available EHR dataset.
• Performed extensive data preprocessing on laboratory measurements, prescription data, and patient demographics.
• Developed and trained machine learning models to predict potential physiological abnormalities across 20 selected measurement items.

Main Results:

• The developed models successfully predicted potential abnormalities in selected physiological measurements.
• Feature importance analysis revealed age, specific active pharmaceutical ingredients, and sex (male/female) as the most influential predictors.
• The models generated predictions in a human-readable sentence format.

Conclusions:

• Machine learning models trained on EHR data can effectively predict physiological effects of multiple drug interactions.
• Patient-specific data, including demographics and drug combinations, are critical for accurate DDI prediction.
• The methodology is adaptable for predicting other physiological measurements and can be applied to different EHR datasets.