PMH34 INTEGRATING COUNTY-LEVEL HEALTH, HUMAN SERVICES, AND CRIMINAL JUSTICE DATA TO PREDICT RISK OF OPIOID OVERDOSE AMONG MEDICAID BENEFICIARIES: A MACHINE LEARNING APPROACH

Abstract

Recent studies indicate the shortcomings of current opioid risk prediction tools and call for developing more advanced models to improve identification of individuals at risk (or no risk) of opioid-related adverse outcomes. We integrated human services data, criminal justice records, and medical examiner’s autopsy data with medical claims data to develop and validate a machine-learning algorithm to predict opioid overdose among Medicaid beneficiaries. Our prognostic study included data of Medicaid beneficiaries in Allegheny County, Pennsylvania (2015-2018). We randomly divided beneficiaries into training, testing, and validation samples, and measured 299 potential predictors in 30-day periods, starting from the first observed claim/record. Gradient boosting machine (GBM) was used to predict a composite outcome (i.e., fatal or nonfatal opioid overdoses) in the subsequent month. We compared prediction performance between the models with medical claims only vs. comprehensive integrated data (i.e., medical claims plus public human services and criminal justice data) using the C-statistic and other metrics (e.g., number needed to evaluate [NNE] to identify one overdose). Beneficiaries were stratified into subgroups by risk-score decile. Beneficiaries in the training (n=79,087), testing (n=79,086), and validation (n=79,086) samples had similar characteristics (age=37.9±18.2 years, female=56.0%, white=48.2%, ≥1 opioid overdose=1.66%). In the validation sample, the GBM algorithm including comprehensive integrated data outperformed the model using medical claims only (C-statistic=0.920; 95%CI=0.914-0.926 vs. C-statistic=0.871; 95%CI=0.863-0.879). In the comprehensive integrated model, individuals in the top risk decile (11.7% of the cohort) had a 0.55% positive predictive value, 99.4% negative predictive value, and NNE of 182. Over 85% of individuals with overdoses were in the top two deciles(n=17,975), having the highest overdose rates (6-55 per 10,000 beneficiaries). Few individuals had overdose episodes in the bottom eight deciles (<1 per 10,000). Machine-learning algorithm integrating comprehensive data appears to improve prediction of opioid overdose risk in a large U.S. county.