Opioid overdose remains a leading cause of preventable death in the United States, with 54,743 opioid‑related deaths in 2024 [1]. The economic and societal burden of opioid use disorder (OUD) and fatal opioid overdose exceeds US $1 trillion annually, including health care expenditures, lost productivity, premature mortality, reduced quality of life, and criminal justice involvement [2].

Although health systems and other stakeholders have instituted multiple interventions to reduce patient risk of overdose [3,4], including promoting safer opioid prescribing and dispensing naloxone, such efforts face 2 noteworthy challenges. First, there are limited and poorly performing tools to help clinicians identify who is truly at risk of overdose, leading to imprecise and burdensome interventions that target an overly broad population [5,6] or fail to identify the high-risk individuals most likely to experience overdose. Second, even with more accurate identification of elevated-risk patients, highly effective strategies to change clinician therapeutic behaviors remain limited.

To address these challenges, our team previously developed and externally validated machine learning (ML) algorithms for overdose prediction. In an application in Medicare data, these algorithms identified more than 90% of overdoses among the top risk groups, whereas commonly used risk classifications such as the high opioid dose measure used by the Centers for Medicare and Medicaid Services missed 70% to 90% of individuals with an actual overdose [3,7,8]. However, even with more accurate risk prediction in clinical settings, prognostic information may not translate into clinical action because clinicians face workflow challenges, biases and clinical inertia, and other barriers [9,10]. Behavioral nudges, which are subtle changes to how choices are presented or how information is framed to influence behavior without restricting choice, can help facilitate action by clinicians [11-16]. One nudge strategy, the accountable justification, prompts clinicians to briefly document a rationale for a clinical decision, and it has increased evidence-based oncology prescribing, reduced unnecessary antibiotic use, and decreased opioid prescribing in prior studies [17-19].

The Machine Learning Prediction and Reducing Overdoses With Electronic Health Record (EHR) Nudges (mPROVEN) study combines more accurate ML risk prediction with behavioral nudges as a targeted intervention to improve prescribing practices associated with overdose risk and, ultimately, reduce overdose risk itself. We integrated a validated ML-driven opioid overdose risk prediction model into a large health system’s EHR to identify patients at elevated risk of opioid overdose. We will conduct a 3-arm pragmatic, cluster randomized controlled trial (RCT) among patients with an elevated predicted risk of overdose, delivering within the EHR risk information, active choice alerts, and accountable justification alerts to primary care providers who see elevated-risk patients.

This study was funded by the US National Institutes of Health in June 2022. Clinical trial enrollment began on March 10, 2025. Enrollment for the primary analysis cohort was completed (n=798) on May 26, 2025. Additional participants continued to be enrolled for secondary analyses (n=1662 as of December 12, 2025). Data analysis for the primary cohort began in January 2026, and results are expected by the middle of 2027.

We describe the protocol for a cluster randomized trial evaluating an ML-guided set of behavioral nudges embedded within the EHR to reduce overdose-related risk in primary care. We anticipate that this intervention will improve evidence-based prescribing behaviors associated with reduced overdose risks among elevated-risk patients and, in turn, reduce overdose risk. By integrating accurate risk prediction with actionable behavioral nudges, this study is expected to provide evidence on a scalable strategy to influence clinicians’ decision-making and improve opioid safety in real-world primary care settings.

Although prior studies have tested various prediction models for overdose [4,34-42] and others have examined behavioral nudging strategies to influence prescriber behavior [15,43-46], we are not aware of any prior randomized trials that test the combined impact of overdose risk prediction and behavioral nudges. Importantly, several overdose risk prediction tools have already entered clinical practice with limited validation and evaluation, underscoring the necessity of rigorous clinical trials to assess their real-world effectiveness and inform policy decisions. By testing these tools in practice in a large RCT, we will provide high-quality evidence on the effectiveness of these interventions, providing clinicians, health system leaders, and policymakers with critical data to guide adoption and implementation.

In designing the trial, we chose practice-level randomization rather than patient- or clinician-level randomization to minimize the risk of contamination across arms and ensure that all clinicians within practices received a consistent intervention. UPMC’s robust clinical informatics team and the presence of a large number of primary care practices operating on a unified EHR platform enabled this practice-level approach, allowing us to isolate the effect of the intervention alone on prescribing practices. Our months-long silent and live testing phases are a strength of the study, enabling calibration and refinement of the alerts to maximize predictive performance while minimizing excessive firing, workflow disruption, and alert fatigue. Although we anticipate that larger practices may generate a higher volume of alerts than smaller practices due to greater patient volume, our stratified randomization by practice size is designed to balance this difference across study arms. Our 3-arm design structure enables assessment of the added benefit of nudges over the risk flag alone.

There are some important potential limitations in this study. First, a challenge in designing trials addressing opioid overdose in primary care is the dynamic nature of how practices respond to the opioid epidemic and how the epidemic itself changes. For example, practices in our usual care arm could begin to adopt safer prescribing behaviors similar to those encouraged by our intervention if UPMC were to introduce parallel system-wide initiatives during the study period; however, we do not anticipate any major changes during our trial period. Other potential challenges include alert fatigue, variable clinician engagement, potential EHR system changes or updates, and model drift over time. We have mitigated these limitations through our iterative nudge design, model fairness and calibration assessments, continuous monitoring through silent and live phases, and proactive adaptation to systematic changes. Second, because overdose events are relatively rare, the trial is not powered to detect differences in overdose outcomes as a primary end point. Instead, we use a composite of prescribing behaviors with strong evidence linking them to reduced overdose risk as a pragmatic and actionable proxy [47]. Third, while we assigned study arm by practice to minimize the risk of contamination, some physicians and other clinicians may work out of multiple departments during the course of the study. Fourth, while our initial plan was for a 6-month intervention period, unexpected changes occurred in Epic within UPMC that required shortening the intervention period for the primary analysis to 4 months. However, we anticipate that most intervention effects will occur within the first four months of the trial, when clinicians receive the initial set of justification alerts. These alerts are programmed to reappear based on predefined criteria. Specifically, certain alerts will reappear 3 months after the index encounter if the underlying risk condition persists (eg, continued high opioid dosage or opioid/benzodiazepine overlap after a prior justification), whereas others reappear sooner if the triggering condition recurs (eg, a new opioid prescription). Regardless, we have preplanned a 6-month secondary outcome to compare any difference in intervention efficacy. Finally, we acknowledge that our intervention will focus on the primary care setting, and while it may generalize to other parts of the health care system that use an EHR, it will not help the many people at risk of overdose who do not engage with the health care system.

Our planned dissemination of study results includes the usual academic channels, including journal publication and presentation at scientific meetings in late 2026 and early 2027. We are also in regular communication with the clinical leadership of primary care in the health system, as well as the major informatics groups, enabling real-time dissemination of findings to interested parties who can implement changes after the clinical trial is completed. Our research team also received supplemental funding from the National Institutes of Health to compile “lessons learned” from our implementation and create a playbook that will guide others in implementing similar overdose prediction tools in the EHR in an ethical and successful manner.

In summary, the mPROVEN study is among the first pragmatic RCTs to integrate ML‑based opioid overdose risk prediction with behavioral nudges directly within a large health system EHR. By combining advances in data science, behavioral economics, and implementation science, the study aims to reduce opioid overdose risk in primary care using a scalable and low-touch intervention to address a high-priority public health issue.