The potential for leveraging machine learning to filter medication alerts.

Area of Science:

• Clinical Informatics
• Artificial Intelligence in Healthcare
• Health Information Technology

Background:

• Medication alerts are crucial for patient safety but can overwhelm clinicians.
• A significant number of medication alerts are ignored, diminishing their effectiveness.
• Predictive modeling offers a potential solution to filter non-critical alerts.

Purpose of the Study:

• To evaluate machine learning (ML) models for predicting ignored medication alerts.
• To develop a system for intelligently filtering alerts to reduce user burden.
• To identify key features influencing user responses to medication alerts.

Main Methods:

• Literature review and expert input to identify predictive features for medication alerts.
• Development and comparison of rule-based and ML models (logistic regression, random forest, support vector machine, neural network, LightGBM).
• Analysis of 3,481,634 medication alerts from University of Utah Health (2019) with a focus on maximizing precision at a sensitivity of 0.99 and false-negative rate <0.01.

Main Results:

• The LightGBM model demonstrated the highest precision (0.192) while meeting the predefined false-negative rate threshold (P < 0.001).
• This ML model has the potential to reduce the overall volume of medication alerts by 54.1%.
• Removing medication order features significantly decreased model precision, highlighting their importance.

Conclusions:

• Machine learning models show significant potential for the intelligent filtering of medication alerts.
• This approach can help mitigate alert fatigue and improve the clinical utility of medication safety systems.
• Further research into feature importance can optimize ML-based alert filtering.