Machine learning models to detect opioid misuse in Emergency Department patients at triage

Area of Science:

• Emergency medicine
• Data science
• Public health

Background:

• Emergency department (ED) encounters are missed opportunities for initiating opioid misuse treatments.
• Manual screening for opioid misuse is resource-intensive and uncommon.
• Machine learning (ML) offers a potential solution for identifying patients at triage.

Purpose of the Study:

• To determine the feasibility of using ML to identify patients with opioid misuse at ED triage.
• To evaluate ML model performance using routinely collected triage data.

Main Methods:

• Retrospective cohort study of 1,123 ED encounters (September 2020 - March 2023).
• Candidate triage features included demographics, acuity, arrival time, chief complaint, comorbidities, and medications.
• Models evaluated using F1 score, AUPRC, accuracy, recall, and AUROC; SHAP for explainability.

Main Results:

• ML models performed comparably to opioid-related diagnosis codes.
• Random Forest (F1=0.75) and Gradient Boosting (F1=0.77) showed high performance.
• Top predictors included prior drug-use diagnosis codes, triage acuity, and age.

Conclusions:

• ML classifiers using triage data are feasible for flagging opioid misuse.
• This approach can enable early harm-reduction interventions.
• Further validation and bias assessment in multi-site studies are recommended.