Classifying Refugee Status Using Common Features in EMR

ABSTRACT Objective Automated and accurate identification of refugees in healthcare databases is a critical first step to investigate healthcare needs of this vulnerable population and improve health disparities. This study developed a machine-learning method, named refugee identification system (RIS) that uses features commonly collected in healthcare databases to classify refugees and non-refugees. Materials and Methods We compiled a curated data set consisting of 103 refugees and 930 non-refugees in Arizona. For each person in the curated data set, we collected age, primary language, and noise-masked home address. We supplemented de-identified individual-level data with state-level refugee resettlement statistics and world language statistics, then performed feature engineering to convert primary language and masked address into quantitative features. Finally, we built a random forest model to classify refugee status. Results Evaluated on holdout testing data, RIS achieved a high classification accuracy of 0.97, specificity of 0.99, sensitivity of 0.85, positive predictive value of 0.88, and negative predictive value of 0.98. The receiver operating characteristic curve had an area under the curve value of 0.98. The source code is available at GitHub ( https://github.com/liliulab/ris ). Discussion and Conclusion RIS is an automated, accurate, and scalable method to predict refugee status. It uses only de-identified information to protect patient privacy. The computational framework is adaptable to address similar challenges in other States. Its application enables large-scale investigation of refugee healthcare needs and improvement of health disparities.