Developing and Evaluating a Machine-Learning Opioid Overdose Prediction & Risk-Stratification Tool in Primary Care (DEMONSTRATE)

Developing and Evaluating a Machine-Learning Opioid Prediction & Risk-Stratification E-Platform (DEMONSTRATE)

This clinical trial aims to evaluate the pilot implementation of a machine-learning (ML)-driven clinical decision support (CDS) tool designed to predict opioid overdose risk within the electronic health record (EHR) system at UF Health Internal Medicine and Family Medicine clinics in Gainesville, Florida. The study will use a pre- versus post-implementation design to compare outcomes within clinics, focusing on measures such as naloxone prescribing rates and opioid overdose occurrences. Researchers will also assess the usability, acceptability, and feasibility of the CDS tool through qualitative interviews with primary care clinicians (PCPs) in the participating clinics.

Study Overview

• Status: Active, not recruiting
• Conditions: Mental Disorders; Chemically-Induced Disorders; Opioid-Related Disorders; Narcotic-Related Disorders; Substance-related Disorders; Opiate Overdose
• Intervention / Treatment: Behavioral: Machine Learning-Based Clinical Decision Support: Overdose Prevention Alert (OPA) Intervention

Detailed Description

This clinical trial evaluates the pilot implementation of a ML-driven CDS tool designed to predict opioid overdose risk within the electronic health record (EHR) system at thirteen UF Health internal medicine and family medicine clinics in Gainesville, Florida.

The implementation process involved backend and frontend development and integration of the CDS tool. For backend integration, the investigators reviewed clinical workflows, designed a data flow plan to incorporate risk scores into patient charts, and collaborated with UF Health IT and Integrated Data Repository (IDR) Research Services to address alert implementation, data flow, server specifications, and responsibilities. Risk assessments approved by UF Health IT and the institutional review board (IRB) ensured secure access to patient health information (PHI) and enabled EHR integration. For frontend development, the investigators used a user-centered design approach to create the CDS tool prototype, incorporating feedback from PCPs during formative interviews to refine the user interface and ensure timely, actionable alerts through the EPIC system without disrupting clinical workflows.

The study primarily aims to assess the usability, acceptance, and feasibility of the CDS tool six months post-implementation through mixed-method evaluations. Researchers will use semi-structured interviews and an online questionnaire to collect feedback from PCPs, focusing on alert usability, preferences, and outcomes. Quantitative analyses will evaluate alert penetration, usage patterns, and PCP actions, while qualitative analyses will explore themes and insights from override comments to guide tool optimization. Researchers will also explore secondary patient-level outcomes using EHR data such as naloxone prescriptions.

• Study Type: Interventional
• Enrollment (Actual): 674
• Phase: Not Applicable

Contacts and Locations

Study Locations

• United States
  • Florida
    • Gainesville, Florida, United States, 32608
      • University of Florida Health Internal Medicine and Family Medicine

Participation Criteria

Eligibility Criteria

• Ages Eligible for Study: Adult; Older Adult
• Accepts Healthy Volunteers: No

Description

Inclusion Criteria:

For PCP level outcomes assessment

• PCPs
• practicing in any of the 13 participating clinics (10 UF Health Family Medicine clinics and 3 UF Health Internal Medicine) in Gainesville, Florida.

For patient level outcomes assessment:

Inclusion criteria: Patients who seen in any of the 9 participating UF Health clinics who

• are aged ≥18 years
• received any opioid prescription in the past year prior to their clinic visit.
• are identified as being at elevated risk for overdose by the ML algorithm. Exclusion Criteria: Patients who
• had malignant cancer diagnosis or hospice care prior to study enrollment

Study Plan

How is the study designed?

Design Details

• Primary Purpose: Health Services Research
• Allocation: N/A
• Interventional Model: Single Group Assignment
• Masking: None (Open Label)

Number of Arms

1

Arms and Interventions

Participant Group / Arm

Intervention / Treatment

Experimental: Overdose Prevention Alert (OPA) Intervention Arm; The intervention arm will receive a ML CDS tool that provides interruptive alerts for patients at elevated risk of opioid overdose, triggered when a clinician signs an opioid order.

Behavioral: Machine Learning-Based Clinical Decision Support: Overdose Prevention Alert (OPA) Intervention; In this study, researchers will pilot test an interruptive, ML CDS tool for opioid overdose risk across thirteen primary care clinics at the UF Health in Gainesville, FL. When a patient is identified by the ML algorithm as having an elevated overdose risk and a PCP signs an opioid prescription for the patient, an Opioid Prevention Alert (OPA) will be triggered. The alert will include the rationale for the patient's elevated risk status and provide three risk mitigation recommendations: optimizing pain treatment and mental health support, reviewing and discussing risks with the patient, and offering naloxone annually if no prior naloxone order is found in the patient's record. PCPs can also select an override reason, such as the patient already has naloxone, declined the intervention, is not present/it is not the right time, or the alert is not relevant/other comments, when appropriate.

What is the study measuring?

Primary Outcome Measures

Outcome Measure	Measure Description	Time Frame
Composite patient-level outcomes related to opioids	The CDS tool will generate an Overdose Prevention Alert (OPA) when a PCP signs an opioid order in Epic®. To evaluate the tool's effectiveness, researchers will conduct within-clinic comparisons (pre- vs. post-implementation) and examine a composite of patient-level outcomes post-implementation, including the proportion of patients having any of the following 6 outcomes: receipt of a naloxone order or prescription fill;; absence of opioid overdose diagnoses and naloxone administration;; absence of ED visits or hospitalizations due to opioid overdose or OUD;; absence of overlapping opioid and benzodiazepine use;; absence of high-dose opioid use (average daily morphine milligram equivalent ≥50);; receipt of referrals to non-pharmacological pain management (e.g., physical therapy, chiropractic care).	From enrollment and up to 12 months (3, 6, 12 months) post implementation of the OPA
PCP's use feedback of the Overdose Prevention Alert (OPA)	An online questionnaire for PCPs who interacted with OPA includes 12 Likert-scale items (4-point scale: 1 = Strongly Disagree to 4 = Strongly Agree) assessing OPA's acceptability, appropriateness, and feasibility: OPA's information was clear.; OPA was easy to use.; OPA helps identify patients at increased overdose risk.; OPA helps understand patient's overdose risk.; OPA provides risk management recommendations.; OPA identifies the right patients with elevated overdose risk.; OPA notifies the correct healthcare team member (i.e., PCPs).; A pop-up alert is an appropriate notification approach.; Signing an opioid order is the right time for OPA.; Alert frequency is appropriate.; I prefer OPA over the legacy naloxone alert (see picture).; I want this OPA to continue to operate in my EHR. Mean scores (with standard deviations [SD]) will be calculated across all items, as well as individual average scores (SD).	From enrollment and up to 7 months post implementation of the OPA

Secondary Outcome Measures

Outcome Measure	Measure Description	Time Frame
Receipt of a naloxone order or prescription fill	Proportion of patients receiving alert who have a naloxone order or prescription fill	From enrollment and up to 12 months (3, 6, 12 months) post implementation of the Overdose Prevention Alert (OPA)
Absence of opioid overdose diagnoses and naloxone administration	Proportion of patients receiving alert who do not have an opioid overdose diagnoses and naloxone administration	From enrollment and up to 12 months (3, 6, 12 months) post implementation
Absence of ED visits or hospitalizations due to opioid overdose or OUD	Proportion of patients receiving alert who do not have ED visits or hospitalizations due to opioid overdose or opioid use disorder (OUD)	From enrollment and up to 12 months (3, 6, 12 months) post implementation
Absence of overlapping opioid and benzodiazepine use	Proportion of patients receiving alert who do not have overlapping opioid and benzodiazepine use	From enrollment and up to 12 months (3, 6, 12 months) post implementation
Absence of high-dose opioid use (average daily morphine milligram equivalent ≥50)	Proportion of patients receiving alert who do not have high-dose opioid use (average daily morphine milligram equivalent ≥50).	From enrollment and up to 12 months (3, 6, 12 months) post implementation
Receipt of referrals to non-pharmacological pain management (e.g., physical therapy, chiropractic care	Proportion of patients receiving alert who have referrals to non-pharmacological pain management (e.g., physical therapy, chiropractic care).	From enrollment and up to 12 months (3, 6, 12 months) post implementation

Collaborators and Investigators

• Sponsor: University of Pittsburgh
• Collaborators: National Institute on Drug Abuse (NIDA); Applied Decision Science
• Investigators: Principal Investigator: Wei-Hsuan Lo-Ciganic, PhD, Division of General Internal Medicine, School of Medicine, University of Pittsburgh, Pittsburgh, PA

Publications and helpful links

General Publications

• Lo-Ciganic WH, Huang JL, Zhang HH, Weiss JC, Kwoh CK, Donohue JM, Gordon AJ, Cochran G, Malone DC, Kuza CC, Gellad WF. Using machine learning to predict risk of incident opioid use disorder among fee-for-service Medicare beneficiaries: A prognostic study. PLoS One. 2020 Jul 17;15(7):e0235981. doi: 10.1371/journal.pone.0235981. eCollection 2020.
• Lo-Ciganic WH, Donohue JM, Yang Q, Huang JL, Chang CY, Weiss JC, Guo J, Zhang HH, Cochran G, Gordon AJ, Malone DC, Kwoh CK, Wilson DL, Kuza CC, Gellad WF. Developing and validating a machine-learning algorithm to predict opioid overdose in Medicaid beneficiaries in two US states: a prognostic modelling study. Lancet Digit Health. 2022 Jun;4(6):e455-e465. doi: 10.1016/S2589-7500(22)00062-0.
• Nguyen K, Wilson DL, Diiulio J, Hall B, Militello L, Gellad WF, Harle CA, Lewis M, Schmidt S, Rosenberg EI, Nelson D, He X, Wu Y, Bian J, Staras SAS, Gordon AJ, Cochran J, Kuza C, Yang S, Lo-Ciganic W. Design and development of a machine-learning-driven opioid overdose risk prediction tool integrated in electronic health records in primary care settings. Bioelectron Med. 2024 Oct 18;10(1):24. doi: 10.1186/s42234-024-00156-3.
• Militello LG, Diiulio J, Wilson DL, Nguyen KA, Harle CA, Gellad W, Lo-Ciganic WH. Using human factors methods to mitigate bias in artificial intelligence-based clinical decision support. J Am Med Inform Assoc. 2025 Feb 1;32(2):398-403. doi: 10.1093/jamia/ocae291.
• Lo-Ciganic WH, Huang JL, Zhang HH, Weiss JC, Wu Y, Kwoh CK, Donohue JM, Cochran G, Gordon AJ, Malone DC, Kuza CC, Gellad WF. Evaluation of Machine-Learning Algorithms for Predicting Opioid Overdose Risk Among Medicare Beneficiaries With Opioid Prescriptions. JAMA Netw Open. 2019 Mar 1;2(3):e190968. doi: 10.1001/jamanetworkopen.2019.0968.
• Hong JJ,Wilson DL,Nguyen K,Gellad WF,Diiulio J,Militello L,Yan S,Harle CA,Nelson D,Rosenberg EI,Schmidt S,Chang CH,Cochran G,Wu Y,Staras SAS,Kuza C,Lo-Ciganic WH

Helpful Links

• Center for Pharmaceutical Policy & Prescribing website

Study record dates

Study Major Dates

• Study Start (Actual): April 8, 2025
• Primary Completion (Estimated): October 7, 2026
• Study Completion (Estimated): October 7, 2026

Study Registration Dates

• First Submitted: January 27, 2025
• First Submitted That Met QC Criteria: January 30, 2025
• First Posted (Actual): February 5, 2025

Study Record Updates

• Last Update Posted (Actual): April 13, 2026
• Last Update Submitted That Met QC Criteria: April 7, 2026
• Last Verified: April 1, 2026

More Information

Terms related to this study

• Keywords: Opioid-Related Disorders; Machine Learning; Decision Support Systems, Clinical; Medical Order Entry Systems; Opiate overdose; Risk Evaluation and Mitigation
• Additional Relevant MeSH Terms: Drug Misuse; Drug Overdose; Prescription Drug Misuse; Opiate Overdose; Narcotic-Related Disorders; Opioid-Related Disorders; Substance-Related Disorders; Mental Disorders; Chemically-Induced Disorders; Investigative Techniques; Methods
• Other Study ID Numbers: STUDY24040038; R01DA050676 (U.S. NIH Grant/Contract)

Plan for Individual participant data (IPD)

• Plan to Share Individual Participant Data (IPD)?: NO
• IPD Plan Description: While transparency and data sharing are critical to advancing clinical research, researchers are unable to share individual participant data (IPD) derived from UF Health EHR due to institutional and legal constraints. The data are governed by strict privacy regulations, including HIPAA, which mandate the protection of patient confidentiality. Additionally, UF Health's policies restrict the dissemination of EHR data to ensure compliance with these regulations and safeguard against the risk of re-identification. As a result, while researchers can report aggregated findings, sharing raw participant-level data is not permissible under current regulatory and institutional frameworks.

Drug and device information, study documents

• Studies a U.S. FDA-regulated drug product: No
• Studies a U.S. FDA-regulated device product: No
• product manufactured in and exported from the U.S.: No