Heterogeneity of risk of substance-related harms in British Columbians – using machine learning and statistical models to predict individual overdose risk with administrative healthcare data

British Columbia is experiencing a public health emergency and toxic drug crisis. Statistical models help us understand this epidemic, and inform treatment and intervention strategies. An assumption of many existing models is that overdose risk is homogeneous throughout the population, ignoring individual risk factors such as past overdoses, treatment history, or other relevant features commonly available in administrative health data. However, key at-risk groups may be particularly affected by this assumption, as impacts may be underestimated. We propose a method of estimating individual overdose risk which leverages available health data. A number of machine learning models will be compared, with a focus on the mechanisms behind predictions, since a greater understanding of factors contributing to overdose risk can assist in targeting populations most in need of intervention and follow-on care. This work will be the first of its kind in BC. It will directly impact policy by providing a method of accounting for heterogeneity of overdose risk in existing models, improving estimation of intervention effects and forecasting, and practice, by the potential of direct application as a clinical tool in the distribution of care.