A machine learning-based clinical predictive tool to identify patients at high risk of medication errors

Area of Science:

• Health Informatics
• Artificial Intelligence in Medicine
• Patient Safety

Background:

• Medication errors pose significant risks to patient morbidity, mortality, and healthcare economics.
• Medication reconciliation is a proven, cost-effective intervention to reduce medication errors by up to 66%.

Purpose of the Study:

• To develop and evaluate a machine learning-based tool for prioritizing patients at risk of medication errors during hospital admission.
• To enhance the efficiency and effectiveness of medication reconciliation by clinical pharmacists.

Main Methods:

• Utilized electronic health records from 7200 patients admitted to Reims University Hospital (2017-2023).
• Trained four machine learning models using 52 variables to predict medication error risk.
• Evaluated model performance using metrics including recall, precision, F1 score, AUROC, and AUCPR.

Main Results:

• The voting classifier model demonstrated strong performance with a recall of 0.75 and an F1 score of 0.70.
• In a retrospective simulation, the voting classifier identified 45% of patients with unintended discrepancies, compared to 21% with the existing tool.
• The machine learning tool showed a 113% improvement over the current method in identifying at-risk patients.

Conclusions:

• Machine learning models can effectively prioritize patients for medication reconciliation upon admission.
• The developed tool significantly enhances the identification of medication errors, improving patient safety.
• This AI-driven approach offers a superior strategy for proactive medication error prevention in hospitals.