A group of researchers has developed a model predicting the risk of opioid overdoses among Medicare beneficiaries that seems to do a better job than efforts using traditional statistical models, perhaps because it leveraged machine learning technology.
Their study, which appears in JAMA Network Open, draws on Medicare claims data. Between 2017 and 2018, researchers studied 560,057 fee-for-service Medicare beneficiaries without cancer who filled one or more prescriptions for opioids between 2011 and 2015.
Every three months, the research team captured data on potential predictors of opioid use, including the patients’ demographics, health status, opioid use patterns and information on the regions where they lived. The researchers then isolated opioid overdoses by scrutinizing inpatient and emergency department claims.
At this point, the researchers attempted to predict the risk of overdose in the three months after participants began opioid treatment using machine learning algorithms. Ultimately, they divided the population into three risk groups by projected risk score, with three-quarters falling into a low-risk group.
The team found that more than 90% of individuals with overdoses had been sorted into the high- and medium-risk groups. In other words, the analysis seems to have done a pretty good job of targeting potential opioid issues.
At this point, a few comments seem in order, particularly given the importance of this topic.
First, it’s interesting to see that this study relied primarily on claims data, which overlaps with but certainly isn’t identical to the data found in inpatient or outpatient medical records.
We’ve reported on at least one other study attempting to predict opioid outcomes, in which investigators seem to have gotten good results out of data found in a hospital EHR. That study analyzed factors such as medical and mental health diagnoses, substance and tobacco use, chronic or acute pain, having received opioid or non-opioid analgesics during the past year and whether they left the hospital with opioid prescriptions to predict whether the patient would end up on chronic opioid therapy.
As I see it, the logical next step for this kind of research is to analyze data related to social determinants of health and use it to predict which patients are at risk for abuse of both prescription and street narcotics. Looking at those factors would be a useful adjunct to data which predicts how position and hospital behavior affects the risk of opioid-related harm.
After all, without SDOH data, providers can do nothing to address other important factors which had an impact on a patient’s risk of opioid abuse, overdose or death. To develop an enlightened and sophisticated population health strategy, and particularly to address problems with drugs of abuse, we truly, truly need to meet patients where they are.
