We extracted diagnosis data for patients in the study cohort from 01–2001 to 06–2018, consisting of 15,750,104 medical encounters, including a patient identifier, service date, diagnosis code, and diagnosis description. We defined phenotypes for OUD using a pre-specified list of International Classification of Diseases (ICD) 9th and 10th edition codes derived from clinical definitions (Table 1). Cases were patients with at least one corresponding ICD code of opioid dependence, opioid abuse or long-term opioid use disorder, as in other claims-based prediction models(Reps et al., 2020). A similar rule set was used to extract alcohol and nicotine use disorders. The control group consisted of patients without any diagnosed major substance use disorder (nicotine, alcohol, opioid, cannabis, cocaine and amphetamine). To determine prescription opioid exposure among controls, we obtained medication records for our study cohort from the Partners Biobank, corresponding to 4,864,561 rows of prescription records, and screened the control group for prescriptions of all major opioids.
