U of A researchers use machine learning to predict future opioid overdose cases

A team of researchers at the University of Alberta has developed a machine learning model that can accurately predict an individual's future risk of opioid overdose using population-level data — a tool that could support earlier interventions, shape public health policy and help save lives.

The study, recently published in Molecular Psychiatry, was led by Bo Cao, associate professor in the Department of Psychiatry, adjunct in Computing Science and Public Health and co-director of the U of A's Computational Psychiatry Group, in collaboration with Alberta Health and experts in computing science, emergency medicine and public health.

The team used anonymized health data from roughly four million Alberta residents to train and test the model. The data, from across multiple years, included physician billing records, hospital visits, emergency department encounters, prescription history and mental health indicators.

Machine learning is a type of computational method used for recognizing patterns in large sets of data.

Using data from 2017, the model built by Cao’s team accurately predicted overdose events in 2018, then was tested on three subsequent years of data to confirm its reliability. It achieved a balanced accuracy of over 80 per cent and identified key predictive factors, including prior treatment for substance use, depression, anxiety disorders and physical injuries such as skin wounds.

While previous models have been based on smaller or more selective groups, such as patients already diagnosed with substance use disorder, Cao’s model analyzes data from the general public regardless of diagnosis status. By relying on objective, routinely collected health system data, the model provides a powerful, updateable tool for assessing overdose risk across a population.

This makes the tool especially promising for policymakers and health administrators, who often lack clear data on where to focus resources or how to intervene early. By flagging individuals at increased risk, the model, with proper consent and deployment, could have the potential to inform education campaigns, clinical outreach and resource allocation, while also helping frontline clinicians.

The research methodology could also be useful in other Canadian provinces with universal health care that collect population-based health administration data. That would need to be verified by future studies.

Opioid overdoses have become a major public health crisis across Canada, where more than 50,000 people have died from opioid-related overdoses since 2016. Opioid-related emergency department visits in Alberta have doubled over recent years, increasing from about 21 per day in 2016 to 43 per day in 2023.

While the research holds promise, the team cautions that it’s only a start and there are limitations to overcome, including how the model accounts for under-reported cases or individuals who do not interact with the health system before an overdose occurs. There is also a non-trivial five to 11 per cent false-positive rate.

The team emphasizes that prediction tools like this should be used ethically and responsibly, especially when dealing with highly stigmatized issues such as substance use.

While the model isn’t yet ready for clinical use, it does offer a powerful proof of concept for how population-level health data and machine learning can work together to address urgent public health challenges.

With further refinement and responsible implementation, this approach could become a valuable tool for guiding prevention strategies, supporting health-care decision-making, and ultimately reducing the number of lives lost to opioid overdose.