Imminent opioid overdose risk prediction using classical and machine learning survival models following first...

Background: Identifying patients at risk of opioid overdose in healthcare settings is critical, yet evidence on predictive models and their performance to predict imminent opioid overdose remains limited.Objective: We compared classical and Machine Learning (ML) survival models to predict 30-day overdose risk following a first opioid-related diagnosis to determine whether algorithmic complexity improves clinical decision support.Methods: We conducted a prospective cohort study using longitudinal Electronic Health Record (EHR) data from the All of Us Research Program (v8; May 2018-October 2023). Adults with first recorded opioid-related diagnosis were followed 30 days for incident overdose. We compared Cox proportional hazards, Weibull Accelerated Failure Time (AFT) (parametric), and four ML survival models (Elastic-Net Cox, Random Survival Forest, Gradient-Boosted Survival Trees, Survival SVM).Results: Among 14,737 individuals, 560 overdoses occurred within 30 days (cumulative incidence 3.8%, 95% CI 3.5-4.1). Prior overdose (aHR 5.77, 95% CI 4.64-7.17) and opioid misuse (aHR 3.72, 95% CI 2.94-4.70) were the strongest predictors and consistently drove risk across models. Test-set C-indices ranged from 0.708 to 0.745, with Weibull AFT and Survival SVM performing best, although performance was similar across models. Calibration was acceptable, with predicted risks closely aligned with observed 30-day risks across low and moderate risk strata, and high-risk groups demonstrated substantially lower overdose-free survival.Conclusions: Patients with prior overdose or opioid misuse diagnosis represent a high-yield target for proactive clinical intervention. Simple, interpretable approaches may be sufficient for identifying high-risk patients, as added algorithmic complexity did not meaningfully improve imminent overdose prediction.