Medical-AI trial: predicting overdose risk with machine learning and nudging safer prescribing in the EHR — curbing opioid overdose (cluster-randomized) (NCT06806163)

Opioid (strong painkiller) overdose is a serious public-health crisis in the United States. Who carries risk depends on many factors, such as history, co-prescriptions, and the course of prescribing, and it is easily missed in a busy clinic. If machine learning (ML) can predict high-risk patients from past data and deliver that to the physician, prescribing might be made safer. This study tests such a mechanism.

Per the registry summary, this cluster-randomized controlled trial tries a clinician-targeted behavioral nudge in the EHR for patients the ML model flags as at elevated risk of opioid overdose. Specifically, it provides a flag in the EHR for high-risk individuals, with and without behavioral nudges (best practice alerts, BPAs), compared with usual care by primary care clinicians.

The main goals are to improve opioid prescribing safety and reduce overdose risk, and the primary outcome is a prescribing-practices composite score. The size is 1,350, and it is recruiting.

What is distinctive is that the design covers not only the AI prediction but how that prediction connects to physician action. A prediction that does not change prescribing is meaningless. Combining a behavioral nudge, a gentle push toward the desired choice rather than a mandate, and comparing flag-alone with flag-plus-nudge gets at the heart of AI implementation: from prediction to action.

That said, this study evaluates effectiveness by prescribing practices and the like; it does not establish a reduction in overdose deaths or the effectiveness of the model. Prediction indicates a level of risk, and individual decisions rest with the physician.