Artificial Intelligence-Enabled Software Prototype to Inform Opioid Pharmacovigilance From Electronic Health Records:...

Area of Science:

• Health Informatics
• Artificial Intelligence in Medicine
• Pharmacovigilance

Background:

• Electronic health records (EHRs) contain valuable patient data for drug safety.
• Artificial intelligence (AI) and natural language processing (NLP) can extract insights from unstructured EHR text.
• Detecting safety signals for US Food and Drug Administration (FDA)-regulated drugs can be enhanced by EHR data.

Purpose of the Study:

• To develop an AI-enabled software prototype for identifying adverse drug event (ADE) safety signals.
• To extract ADE signals from free-text discharge summaries in EHRs.
• To improve opioid drug safety and support FDA research activities.

Main Methods:

• Developed a web-based software prototype using keyword/trigger-phrase searching, rule-based algorithms, and deep learning.
• Utilized MedSpacy for section identification and Spark NLP for Healthcare for named entity recognition in discharge summaries.
• Extracted candidate ADEs for opioid drugs from the MIMIC III database and gathered feedback from 15 FDA staff members.

Main Results:

• Successfully identified known, opioid-related adverse drug reactions from EHR text.
• Achieved AI model performance metrics: accuracy (0.66), recall (0.69), precision (0.64), and F1-score (0.67).
• FDA participants found the prototype highly desirable for usability, visualizations, and potential to support drug safety signal detection, saving time and manual effort.

Conclusions:

• The novel prototype automates the extraction and analysis of clinically useful information from unstructured EHR text using AI.
• It enhances efficiency in using real-world data for opioid drug safety monitoring.
• The tool increases data usability for regulatory review and reduces the manual research burden.