The Daily Newsletter
Far more people in the US may have died from opioids in the past two decades than previously reported, according to a new analysis of unclassified drug deaths carried out using machine-learning algorithms.
Elaine Hill and her colleagues at the University of Rochester, New York, were examining data on drug overdose deaths when they realised that 22 per cent of such cases reported between 1999 and 2016 were listed on death certificates as overdoses without specifying the substance involved.
“We found that remarkable, given the scale of the issue,” says team member Andrew Boslett.
The team tried to estimate what percentage of these unclassified deaths were due to opioids by analysing the coroners’ and medical reports from opioid overdoses and unclassified overdoses.
First, the researchers used machine-learning algorithms to analyse deaths that had been recorded as being due to opioid overdose. They were able to identify common factors that could signify the involvement of opioids, such as descriptions of long-term pain and arthritis.
Using this information, the team estimates that 72 per cent of unclassified overdose deaths involved opioids. This finding suggests that 99,160 more people in the US have died from opioid overdoses than previously thought, an underestimate of 28 per cent. According to these new results, a total of more than 450,000 people in the US have died from an opioid overdose since 1999.
“We were initially surprised by this data, but then it felt plausible when looking at work that’s been done on a local level,” says Hill.
According to the analysis, some states under-reported opioid overdose deaths far more than others. Pennsylvania and Delaware performed the worst in this regard.
The team is now studying exactly why so many US overdose deaths have been, and continue to be, unclassified.
“It’s well known in the field that opioid-related overdose deaths are generally undercounted,” says Jennifer McNeely at New York University. She says that only by knowing where the epidemic is having the biggest impact can we best target interventions.