Extraction and validation of patient housing and food insecurity status in a large electronic health records database using selective prediction and active learning

ObjectiveInformation on patient social determinants of health is frequently recorded in unstructured clinical notes, making it inaccessible for researchers and policymakers. We aimed to extract and validate food and housing insecurity status on a large electronic health record-derived patient cohort by combining selective prediction and active learning. Materials and MethodsManually labeled charts selected via active learning were used to train L1-regularized logistic regression models to identify the presence of food insecurity (N=372, 42% event rate) and housing insecurity (N=559, 36% event rate) in clinical notes. In addition to validating predictions against labeled data, we further validated predictions on an additional unlabeled dataset through associative studies with demographic, clinical, and environmental variables with known associations with food and housing insecurity. ResultsThe food insecurity model had AUC=0.83, sensitivity=0.90, PPV=0.90, and undetermined rate=0.59 (n=149); the housing insecurity model had AUC=0.81, sensitivity=0.50, PPV=1, and undetermined rate=0.65 (n=224). Out of 4,337 unlabeled patients, the 395 (9%) patients predicted to have food insecurity were more likely to be Hispanic/Latino (48% vs 24%, p<0.001) and have diabetes (34% vs 12%), hypertension (43% vs 11%), and heart disease (12% vs 0.7%) (p<0.001 for all). DiscussionSelective prediction and active learning can facilitate efficient labeling of social determinants of health from unstructured EHR data to identify vulnerable populations and targets for healthcare system and policy intervention. ConclusionMachine learning can be used to extract high-fidelity information on patient food and housing insecurity status.