In 2021, there were more than 100,000 overdose deaths in the US, many of which involved synthetic opioids such as illicitly manufactured fentanyl and fentanyl analogs [1]. In the US, at least half of people incarcerated at any given time meet diagnostic criteria for a substance use disorder [2] and more than 20% have opioid use disorder (OUD) [3]. Additionally, at least 20% of people with OUD have been involved in the criminal legal system [4]. After release from incarceration, people are at particularly high risk for opioid overdose due to a complex combination of reduced tolerance, stress and anxiety, lack of social support, and struggling to meet basic needs such as housing [5], [6].

There are large racial disparities in incarceration. Black non-Hispanic individuals are twice as likely to be arrested for drug-related violations than White non-Hispanic individuals, despite similar patterns of drug use [7]. However, In Massachusetts jails between 2015–2020, the estimated rate of overdose deaths per 100,000 people was highest among White individuals, when compared to Black and Latinx individuals [8].

People who have been incarcerated have up to 40 times the risk of an opioid overdose death at two-weeks post-release compared to the non-incarcerated population, with this risk remaining high for several months [9]. Carceral systems rarely screen all individuals for OUD or provide them with medications for opioid use disorder (MOUD) [10], leaving people who return to substance use post-release more vulnerable to overdose during community re-entry. In addition, several structural and social factors increase an individual’s risk of opioid overdose post-release, such as living in a low socioeconomic neighborhood [11], [12].

It is critical to identify individuals most at risk of overdose, but the lack of available administrative data or data linkage infrastructure makes identifying these individuals difficult [13]. Some demographic and incarceration-specific predictors of overdose following release from incarceration have been proposed in prior literature, such as age, sex, race, length of incarceration, security level, and type of conviction [12]. Machine learning methods can be useful for identifying predictors of opioid overdose in this population. While machine learning methods have been used for predicting the risk of opioid overdose with electronic health records [14], [15], [16], these analyses have not been extended to incarcerated individuals. Decision trees are a form of supervised machine learning model that are more flexible than traditional statistical methods, such as logistic regression, but still maintain interpretability [17]. Our objective was to use decision trees to identify individual, social, and structural factors related to opioid overdose in the 90 days following release from Massachusetts prisons for the population overall and stratified by race/ethnicity, as criminal legal involvement may differentially affect the risk of opioid overdose by race/ethnicity. Additionally, we aimed to evaluate the predictive performance of these models.