Predicting and differentiating opioid and non-opioid drug poisonings using health records data.

INTRODUCTION: To facilitate effective targeting of overdose prevention programs, this research developed and evaluated prediction models to identify people at highest risk of specific types of drug overdose or poisoning. Using samples of mental health specialty visits and general medical visits with mental health or substance use diagnoses, this research examined how well prediction models using health records data perform in predicting either any drug poisoning or opioid-involved poisoning and how the specific predictors of opioid-involved poisoning differ from general predictors of any drug poisoning. METHODS: Records data regarding mental health and general medical visits between 2015 and 2019 in four large health systems were used to develop two-step models predicting any poisoning or overdose and differentiating opioid-involved poisonings from other poisonings. Random forest models were developed in random samples of 70 % of visits and validated in held out 30 % samples. RESULTS: Among 19,130,028 visits, 114,911 were followed by a poisoning event and 12,758 by a poisoning involving opioids. A first-step model predicting any poisoning had moderate accuracy, with AUCs of 0.778 among mental health specialty visits and 0.767 among general medical visits. The two-step model to specifically predict opioid involved poisoning had superior performance (AUCs = 0.895 among mental health specialty visits and 0.915 among general medical visits). Predictors of any poisoning included prior self-harm, accidental poisoning, and mental health service use. Specific predictors of opioid-involved poisoning included prior substance use disorder diagnoses and opioid dispensings. LIMITATIONS: Health records data would not identify overdoses or poisonings that do not present for health care and would not identify use of non-prescribed opioids. Finding may not generalize to settings with different patterns of diagnosis or service use. CONCLUSIONS: Models to predict any overdose or poisoning from health records data in four large health systems had only moderate accuracy. Patients at highest risk for opioid poisoning can be more accurately identified by prediction models specifically focused on opioid poisoning, most strongly influenced by indicators of substance use disorder.