Jones et al. explored whether routine clinical and administrative data could be used to flag short-term risk of intentional self-harm or suicide among people receiving opioid agonist treatment (OAT). The team used a retrospective cohort of 46,330 individuals in New South Wales, along with linking hospital, emergency department, mental-health, incarceration, and mortality records. The researchers built machine-learning models to predict events within a 30-day window. They found gradient boosting performed best as it reached an AUC of 0.82, which allowed them to identify 30 major predictors. The strongest signals included recent emergency department visits, hospitalizations for intentional overdose or poisoning, and encounters involving borderline personality disorder, substance dependence, or depression/anxiety. A recent release from incarceration also increased risk. When applied to 2017 data, the model identified roughly 18% of OAT patients as high-risk at a given threshold. It captured about 69% of all self-harm or suicide events. The study shows that short-term risk is concentrated among individuals with frequent acute-care contact and complex psychiatric comorbidities. Furthermore, machine-learning models can outperform simpler rule-based approaches.