Related Experiment Video
Updated: Sep 27, 2026
High-throughput and Comprehensive Drug Surveillance Using Multisegment Injection-Capillary Electrophoresis-Mass Spectrometry
Published on: April 23, 2019
Identifying documented active injection drug use among adults presenting to an urban emergency department with
Ahmed Thiam1, Gerardo Martínez2, Sarah Larney3
1École de santé publique, Université de Montréal, C.P. 6128, succursale Centre-Ville, Montréal, Québec, H3C 3J7, Canada; Centre de recherche en Santé Publique, Université de Montréal and CIUSSS du Centre-Sud-de-l'Île-de-Montréal, Pavillon 7101 avenue du Parc, 3e étage, CP 6128, Succursale Centre-Ville, Montréal, Québec, H3C 3J7, Canada; Centre de recherche du Centre hospitalier de l'Université de Montréal, 900 Rue Saint-Denis, Montreal, Quebec, H2X 0A9, Canada.
Background:
People who inject drugs (PWID) frequently present to emergency departments (ED) with bacterial infections. Identifying this population in electronic health data is challenging because injection practices are not explicitly coded. Most existing algorithms were developed in inpatient or population-based settings and have not been validated for ED data.
Methods:
The source cohort included 27,289 ED episodes of care for adults evaluated at an urban tertiary hospital between 2012 and 2022 for skin and soft tissue infections, osteoarticular infections, or infectious endocarditis potentially related to injection drug use (IDU). A random sample of 4000 distinct patients was selected for manual chart review to establish PWID status during the index episode. Four classifiers (logistic regression, Random Forest, AdaBoost, and XGBoost) were trained using nested cross-validation on 75% of the sample. The final model was selected based on mean F1 score across the outer folds and evaluated on a held-out 25% test set.
Results:
Active PWID represented 5.6% of the sample. Random Forest had the highest mean F1 across outer folds (0.70) and was selected as the final model. On the test set, this model achieved sensitivity of 0.80, specificity of 0.96, positive predictive value of 0.51, negative predictive value of 0.99, and an F1 score of 0.62.
Discussion:
Documented active IDU can be identified with good performance using routinely collected electronic health data from the ED. Such algorithms can support epidemiologic monitoring and service planning for injection-related infections, but require external validation before broader implementation.
