Prediction-based threshold for medication alert.

Area of Science:

• Clinical Informatics
• Pharmacovigilance
• Machine Learning in Healthcare

Background:

• Computerized physician order entry (CPOE) systems often use static thresholds for medication alerts.
• Static thresholds can lead to alert fatigue and may miss excessive medication doses below the set limit.

Purpose of the Study:

• To develop and evaluate a prediction-based approach for setting medication alert thresholds.
• To reduce physician alert fatigue and improve the detection of potentially excessive medication orders.

Main Methods:

• Utilized a random forest algorithm to build a predictive model for medication doses.
• Employed boxplots to determine dynamic thresholds based on prediction results.
• Evaluated the approach on eight high-alert drugs in a hospital setting.

Main Results:

• The prediction model demonstrated high performance for most drugs, with exceptions for two.
• Prediction-based thresholds reduced the number of alerts by approximately 50% compared to static thresholds.
• The new method identified medication cases missed by static thresholds.

Conclusions:

• Prediction-based thresholds offer a practical advantage over static thresholds in CPOE systems.
• This approach leverages collective physician experience to enhance medication safety.
• Dynamic thresholds can optimize alert systems, balancing sensitivity and specificity for better patient care.