Identifying and classifying opioid-related overdoses: A validation study.

Area of Science:

• Health Informatics
• Public Health
• Data Science

Background:

• Opioid overdose is a major public health crisis.
• Accurate identification and classification of overdoses are crucial for intervention and research.
• Existing methods using coded data have limitations in capturing the full scope of overdoses.

Purpose of the Study:

• To develop and validate algorithms for identifying and classifying opioid overdoses.
• To leverage both coded data and clinical text from electronic health records (EHRs) using natural language processing (NLP).
• To assess algorithm performance and portability across different healthcare systems.

Main Methods:

• Utilized data from an integrated healthcare system (Kaiser Permanente Northwest, 2008-2014).
• Included International Classification of Diseases (ICD-9/10) codes, clinical notes, and prescription records.
• Assessed algorithm performance (sensitivity, specificity, PPV, NPV) against medical chart review and tested portability in other systems.

Main Results:

• Code-based algorithms showed excellent performance for opioid-related overdoses (97.2% sensitivity, 84.6% specificity) and heroin-involved overdoses (91.8% sensitivity, 99.0% specificity).
• NLP-enhanced algorithms improved classification accuracy for suicide/suicide attempts and substance abuse-involved overdoses.
• The opioid overdose algorithm demonstrated strong portability across different healthcare settings.

Conclusions:

• Code-based algorithms are effective for detecting opioid-related overdoses and classifying heroin involvement.
• NLP-enhanced algorithms offer significant improvements for classifying complex overdose types like suicides/attempts and substance abuse.
• These validated algorithms, especially NLP-enhanced versions, are valuable tools for research in healthcare settings with NLP capabilities.