Population-level individualized prospective prediction of opioid overdose using machine learning.

Area of Science:

• Public Health
• Data Science
• Epidemiology

Background:

• The North American opioid overdose epidemic has worsened, particularly during the COVID-19 pandemic.
• Prospective, population-level prediction of opioid overdose (OpOD) has been lacking.
• Existing studies have not utilized machine learning for individualized, prospective OpOD risk assessment at a population scale.

Purpose of the Study:

• To develop and validate a machine learning (ML) model for individualized, prospective prediction of opioid overdose (OpOD) at a population level.
• To utilize de-identified provincial administrative health data for OpOD risk prediction.
• To assess the model's performance in predicting OpOD cases across multiple years.

Main Methods:

• A cohort of approximately 4 million individuals was used to train an ML-based OpOD prediction model.
• The model was validated on data from subsequent years (2018-2020) to predict OpOD cases in 2019-2021.
• Predictive performance was evaluated using metrics such as balanced accuracy, sensitivity, specificity, and AUC.

Main Results:

• The ML model achieved high balanced accuracy rates: 83.7% (2018), 81.6% (2019), and 85.0% (2021).
• Key predictors for OpOD included healthcare utilization for substance use, depression, anxiety disorders, and superficial skin injuries.
• Leading predictors were identified from Canadian Institute for Health Information (CIHI) data and physician billing claims.

Conclusions:

• Machine learning enables accurate, individualized prediction of future opioid overdose (OpOD) cases using existing population-level health data.
• The developed model demonstrates potential for informing targeted public health interventions and policy planning to address the opioid crisis.
• This approach offers a novel method for proactive identification of individuals at high risk for opioid overdose.