The AI modeling method called ensemble learning has proven capable of predicting opioid overdoses within a month of patients’ receiving prescriptions of these drugs.
The approach merges predictions from multiple discrete algorithms. In the study testing it for this application, ensemble learning was able to forecast both lethal and nonlethal risks.
The study’s authors suggest such predictions could be aggregated to guide overdose prevention at the local, county and regional levels.
The technique was developed by researchers at Vanderbilt University Medical Center in Nashville and the Tennessee Department of Health. It’s described in a paper published Oct. 19 in the Journal of the American Medical Informatics Association.
Vanderbilt application developer Michael Ripperger, psychiatrist and biomedical informaticist Colin Walsh, MD, and colleagues trained, ensembled and calibrated 10 models on data from more than 3 million patients who were collectively given more than 71 million painkiller prescriptions across Tennessee from 2012 to 2017.
They found that close to 60% of all overdoses—2,574 fatal and 8,455 nonfatal—occurred within 30 days of the corresponding prescription’s filling.
As for accuracy, the performance of both the fatal and nonfatal models improved with ensembling, the authors report.
They note that risk concentration analyses showed two ensembling methods in the test set concentrated 47% to 52% of the overdose outcomes within the top quantiles of predicted probabilities. Moreover, both top quantiles contained 10% of the test set predictions.
The team found one of the most reliably predictive factors for both fatal and nonfatal cases was the total quantity of prescriptions.
“Notably, overlapping benzodiazepine prescriptions were more important in the prediction of fatal opioid-related overdose than nonfatal,” the authors comment. “Multidrug combinations have been known to play a large role in the fatality potential of opioid-related overdoses and benzodiazepines have a synergistic respiratory depressant effect when taken with opioids.”
Ripperger and colleagues stress that neither the fatal nor the nonfatal prediction tool is appropriate for clinical use.
“While it is possible that clinically actionable subgroups may exist within the high-risk tiers, given the size of this study, most localized clinical interventions would likely see highly variable calculated individual risk and unacceptably high false positives,” they comment.
Current actionability of these models rests upon their ability to ascribe relative risk geographically within Tennessee. Studies of their ability to predict counties and regions at highest risk in need of public health resource allocation are underway. Overdose prevention is currently directed after harm has already occurred—for example, basing ‘high impact area’ designations on deaths that have already occurred, not those we seek to prevent.”
The researchers obtained their large datasets from statewide sources, including a prescription monitoring program, hospital discharge information system and Tennessee vital records.
They note current overdose prevention in the Volunteer State relies on anecdotal descriptions and retrospective analyses, neither of which supplies prognoses.
In coverage by Vanderbilt’s news operation, study co-author Ben Tyndall, PhD, of the Tennessee Department of Health says the agency is now working toward “more granular risk quantification to improve prevention targeting for underserved populations.”