Identifying emergency department patients at high risk for opioid overdose using natural language processing and...

Area of Science:

• Data Science
• Public Health
• Clinical Informatics

Background:

• Emergency departments (ED) are critical for identifying individuals at high risk of opioid overdose.
• This study aimed to develop machine learning (ML) models to predict fatal opioid overdose within 12 months post-ED visit.

Purpose of the Study:

• To predict opioid overdose death risk in patients following an emergency department visit.
• To leverage electronic health records (EHR), including clinical notes, for predictive modeling.

Main Methods:

• Merged EHR data with opioid overdose mortality records (2011-2019).
• Utilized mutual information for feature selection, reducing 1336 features to 50.
• Trained and validated XGBoost, random forest, and regression models on 70% and 30% of the sample, respectively.

Main Results:

• Feature selection identified 37 significant predictors from EHR clinical notes.
• Models achieved high accuracy (92%), precision (75%), and recall (57%) in predicting opioid overdose death.
• All models demonstrated satisfactory calibration with a >0.5 probability threshold.

Conclusions:

• ML algorithms using structured and unstructured EHR data effectively identify patients at risk of fatal opioid overdose.
• These predictive tools can guide interventions for at-risk patients, improving clinical decision-making.
• Developed models enhance the timeliness and effectiveness of ED-initiated services for opioid use disorders.