Implementation of an opioid use disorder (OUD) machine-learning phenotype in real-time for the ADAPT clinical trial
4 affiliations
- 1Department of Emergency Medicine, Yale University School of Medicine, New Haven, Connecticut, United States of America
- 2Biomedical Informatics and Data Science, Yale University School of Medicine, New Haven, Connecticut, United States of America
- 3Department of Biostatistics, Yale School of Public Health, New Haven, Connecticut, United States of America
- 4Department of Emergency Medicine, University of Virginia, Charlottesville, Virginia, United States of America
Abstract
We developed and deploy a real‑time, electronic health record‑integrated machine learning phenotype to identify emergency department patients with opioid use disorder for prospective clinical trial screening and buprenorphine initiation. We conducted a multi‑phase study across three emergency departments in a single United States health system from 2014 to 2025. Using visit‑level data available at or before triage, we trained a random‑forest classifier to estimate opioid use disorder risk and embedded scoring in the electronic health record to trigger point‑of‑care alerts for trial eligibility review. A high-specificity computable training label supported retrospective development, and a clinician gold-standard reference was established through structured chart review. Performance was summarized using discrimination, calibration, and threshold-based classification metrics; prospective validation used a stratified random sample of flagged and unflagged encounters. Retrospective discrimination compared to the high-specificity computable training label was high (area under the receiver operating characteristic curve, 0.99; 95% CI, 0.98–0.99; area under the precision-recall curve, 0.92; 95% CI, 0.90–0.94), and calibration plots informed operating‑point selection for real‑time use. In prospective gold‑standard validation (n = 217), 89.3% of model-positive encounters and 95.7% of model-negative encounters agreed with physician adjudication at the prespecified threshold. Because the sample was stratified on model prediction, these are within-stratum agreement estimates; weighting them by the source-population flag rate (28,284 of 866,569 eligible encounters, 3.26%) yields design-weighted estimates of population-level prevalence (7.1%), sensitivity (0.40, 95% CI 0.24-0.64), and specificity (0.996). An electronic health record‑embedded, machine learning phenotype can accurately and feasibly identify emergency department patients with opioid use disorder in real time, streamlining clinical trial enrollment and treatment initiation. Ongoing work will report operational metrics (e.g., alert volume and latency), monitor performance drift and equity across subgroups, and evaluate downstream clinical and trial outcomes. Author summary: In this study, we wanted to find a better way to identify people visiting the emergency department who may be living with opioid use problems. These patients are often hard to recognize quickly, even though timely support, including offering effective medications, can greatly improve their health and safety. Traditional approaches rely heavily on clinicians noticing certain patterns in real time, which can be challenging during busy emergency care. To address this gap, we examined over a decade of information from a large health system and developed a clinical decision support tool that reviews routine information collected at the start of an emergency visit. The tool gives clinicians a simple, real-time signal when a patient may benefit from additional evaluation or treatment. Our early experience shows that this approach can help teams identify patients more consistently and connect them to care more quickly. We hope this work will make it easier for hospitals to support people affected by opioid use and to run clinical studies that improve future treatment options.
Related
- An explainable machine learning model based on multidimensional task-state electroencephalography for identifying attention-deficit/hyperactivity disorder comorbid developmental dyslexia
- A Curated Phenotype-Linked HIV-1 Sequence Dataset for Machine Learning Approaches to Tropism Prediction, Disease Progression and Proviral Reservoir Studies.
- Deciphering Vascular Smooth Muscle Cell Phenotype Switching via Integrated Single-Cell Transcriptomics and Machine Learning: CX3CR1 and COL1A1 as Novel Diagnostic Biomarkers for Atherosclerosis
- Window leakage in clinical EEG machine learning: matched subject-independent evaluation across ADHD and Parkinson cohorts
- Sensory-sleep dysfunction is a shared phenotype across genetically distinct neurodevelopmental disorder models
BibTeX
@article{li2026implementation,
title = {{Implementation of an opioid use disorder (OUD) machine-learning phenotype in real-time for the ADAPT clinical trial}},
author = {Huan Li and Mark Iscoe and John Lutz and Carolina Diniz Hopper and Sabrina Fried and Josue Minaya and Caroline Raymond King and Olga Reykhart and Hyung Paek and Daniella Meeker and Edward R. Melnick and R. Andrew Taylor},
journal = {PLOS Digital Health},
year = {2026},
month = sep,
doi = {10.1371/journal.pdig.0001140},
url = {https://journals.plos.org/digitalhealth/article?id=10.1371/journal.pdig.0001140},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - Implementation of an opioid use disorder (OUD) machine-learning phenotype in real-time for the ADAPT clinical trial
AU - Huan Li
AU - Mark Iscoe
AU - John Lutz
AU - Carolina Diniz Hopper
AU - Sabrina Fried
AU - Josue Minaya
AU - Caroline Raymond King
AU - Olga Reykhart
AU - Hyung Paek
AU - Daniella Meeker
AU - Edward R. Melnick
AU - R. Andrew Taylor
PY - 2026
DA - 2026/09/03
JO - PLOS Digital Health
DO - 10.1371/journal.pdig.0001140
UR - https://journals.plos.org/digitalhealth/article?id=10.1371/journal.pdig.0001140
AB - We developed and deploy a real‑time, electronic health record‑integrated machine learning phenotype to identify emergency department patients with opioid use disorder for prospective clinical trial screening and buprenorphine initiation. We conducted a multi‑phase study across three emergency departments in a single United States health system from 2014 to 2025. Using visit‑level data available at or before triage, we trained a random‑forest classifier to estimate opioid use disorder risk and embedded scoring in the electronic health record to trigger point‑of‑care alerts for trial eligibility review. A high-specificity computable training label supported retrospective development, and a clinician gold-standard reference was established through structured chart review. Performance was summarized using discrimination, calibration, and threshold-based classification metrics; prospective validation used a stratified random sample of flagged and unflagged encounters. Retrospective discrimination compared to the high-specificity computable training label was high (area under the receiver operating characteristic curve, 0.99; 95% CI, 0.98–0.99; area under the precision-recall curve, 0.92; 95% CI, 0.90–0.94), and calibration plots informed operating‑point selection for real‑time use. In prospective gold‑standard validation (n = 217), 89.3% of model-positive encounters and 95.7% of model-negative encounters agreed with physician adjudication at the prespecified threshold. Because the sample was stratified on model prediction, these are within-stratum agreement estimates; weighting them by the source-population flag rate (28,284 of 866,569 eligible encounters, 3.26%) yields design-weighted estimates of population-level prevalence (7.1%), sensitivity (0.40, 95% CI 0.24-0.64), and specificity (0.996). An electronic health record‑embedded, machine learning phenotype can accurately and feasibly identify emergency department patients with opioid use disorder in real time, streamlining clinical trial enrollment and treatment initiation. Ongoing work will report operational metrics (e.g., alert volume and latency), monitor performance drift and equity across subgroups, and evaluate downstream clinical and trial outcomes. Author summary: In this study, we wanted to find a better way to identify people visiting the emergency department who may be living with opioid use problems. These patients are often hard to recognize quickly, even though timely support, including offering effective medications, can greatly improve their health and safety. Traditional approaches rely heavily on clinicians noticing certain patterns in real time, which can be challenging during busy emergency care. To address this gap, we examined over a decade of information from a large health system and developed a clinical decision support tool that reviews routine information collected at the start of an emergency visit. The tool gives clinicians a simple, real-time signal when a patient may benefit from additional evaluation or treatment. Our early experience shows that this approach can help teams identify patients more consistently and connect them to care more quickly. We hope this work will make it easier for hospitals to support people affected by opioid use and to run clinical studies that improve future treatment options.
SN - 2767-3170
ER -