New machine-learning model found effective at predicting risk of opioid overdose
Traditional statistical approaches to identifying people at high risk for opioid overdose misclassify and target many who are not truly at high risk. Now researchers who studied Medicare beneficiaries who have at least one opioid prescription have developed a way to use machine learning to more effectively predict the risk of an overdose.
An alternative analytic approach, machine learning allows scientists to assess complex interactions of large data that can reveal hidden patterns and generate more accurate predictions in clinical settings.
The study, by researchers at the University of Florida, University of Pittsburgh, Carnegie Mellon University and the University of Utah, appears in JAMA Network Open.
“The ability to identify such risk groups has important implications for policymakers and insurers who currently target interventions based on less-accurate measures to identify patients at high risk,” said Wei-Hsuan “Jenny” Lo-Ciganic, Ph.D., M.S., M.S. Pharm., an assistant professor of pharmaceutical outcomes and policy at the UF College of Pharmacy, who was the lead author on the study.
“Our model was effective in dividing the participants into three risk groups according to predicted risk score, with three-quarters in a low-risk group with a negligible rate of overdose, and more than 90 percent of individuals who overdose captured in the high- and medium-risk groups,” said Jeremy C. Weiss, M.D., Ph.D., an assistant professor of health informatics at Carnegie Mellon University’s Heinz College who participated in the study.
In 2017-2018, researchers studied 560,057 fee-for-service Medicare beneficiaries without cancer who filled one or more prescriptions for opioids between 2011 and 2015. Those individuals were randomly assigned and equally separated into training, testing and validation samples. Every three months, the researchers measured potential predictors of opioid overdose, including participants’ socio-demographic characteristics (e.g., age, sex, disability status), health status, patterns of opioid use, and factors related to the participants’ practitioners and the regions where they lived. The researchers then identified opioid overdoses from inpatient and emergency room insurance claims and predicted the risk of overdose in the three months after participants began treatment with the drugs.
The study found machine-learning algorithms performed well at predicting risk of opioid overdose and at identifying subgroups of patients at similar risk of overdose, especially when it came to identifying individuals with a low risk of overdose. On the basis of their findings, the researchers concluded their approach outperformed other methods for identifying risk, such as traditional statistical models.
The authors caution that their study only looked at patients who obtained opioids from medical settings, not those who received them from nonmedical settings, which are not captured in claims data. In addition, the study looked at overdoses in medical settings, not overdoses outside those settings.
“Machine-learning models that use administrative data appear to be a valuable and feasible tool for identifying more accurately and efficiently individuals at high risk of opioid overdose,” said Walid Gellad, M.D., M.P.H., an associate professor of medicine at the University of Pittsburgh and senior author on the study. “Although they are not perfect, these models allow interventions to be targeted to the small number of individuals who are at much greater risk.”
The study was funded by the National Institute on Drug Abuse and supported in part by the Pharmaceutical and Manufacturers of America Foundation.