Abstract
BACKGROUND AND AIMS: People with opioid use disorder are at increased risk of intentional self-harm and suicide. Although risk factors are well known, most tools for identifying individuals at highest risk of these behaviours have limited clinical value. We aimed to develop and internally validate models to predict intentional self-harm and suicide risk among people who have been in opioid agonist treatment (OAT). DESIGN: Retrospective observational cohort study using linked administrative data. SETTING: New South Wales, Australia. PARTICIPANTS: 46 330 people prescribed OAT between January 2005 and November 2017. MEASUREMENTS: Intentional self-harm and suicide prediction within a 30-day window using linked population datasets for OAT, hospitalisation, mental health care, incarceration and mortality. Machine learning algorithms, including neural networks and gradient boosting, assessed over 80 factors during the last 3, 6 and 12 months. Feature visualisation using SHapley Additive exPlanations. FINDINGS: Gradient boosting identified 30 important factors in predicting self-harm and/or suicide. These included the most recent frequency of emergency department presentations; hospital admissions involving mental disorders such as borderline personality, substance dependence, psychosis and depression/anxiety; and recent release from incarceration. The best fitting model had a Gini coefficient of 0.65 [area under the curve (AUC) = 0.82] and was applied to 2017 data to estimate the probability of self-harm and/or suicide. On average 46 people (0.16%) (from a total of 28 000 people in OAT) experienced intentional self-harm or suicide per month. Applying a 0.15% probability threshold, approximately 5167 people were classified as high risk, identifying 69% of all self-harm or suicide cases per month. This figure reduced to 450 per month after excluding people already identified in the previous month. CONCLUSIONS: Among people in opioid agonist treatment, administrative linked data can be used with advanced machine learning algorithms to predict self-harm and/or suicide in a 30-day prediction window.