Partners involved: Pfizer, Regenstrief Institute, and Eskenazi Health
The Question
AFib (short for atrial fibrillation), a common heart rhythm disorder in adults, can have disastrous consequences including life-threatening blood clots and stroke if left undetected or untreated. In the United States, AFib affects an estimated 12 million adults, with approximately 11 percent remaining undiagnosed and up to 23 percent over a two-year period, despite the availability of effective preventive therapies.
While predictive models for AFib risk exist, most have been evaluated retrospectively or outside of routine clinical workflows.
The question driving this work was: Can Regenstrief build an AFib risk model from real-world data, validate and carry it into clinical workflows to generate implementation-informed evidence relevant to prevention, management and downstream outcomes.
The Analysis
Why existing evidence was insufficient
Most industry collaborations engage a research partner at one stage of the lifecycle – for retrospective validation, or for a pilot study, but rarely for both. Pfizer, the Regenstrief Institute and Eskenazi Health, a comprehensive health system, structured this uniquely.
UNAFIED is a highly accurate artificial intelligence (AI) prediction model which uses machine learning to parse information acquired from a patient’s electronic health record (EHR) to predict whether a patient has or might develop detectable AFib within the following two years.
UNAFIED used patient data from the Indiana Network for Patient Care (INPC), a 30+ year, multi-institutional health information exchange spanning more than 150 contributing organizations across Indiana. UNAFIED is an acronym for Undiagnosed Atrial Fibrillation prediction using electronic health data.
That same team validated the model in a separate INPC cohort, conducted a proof-of-concept implementation within a production Epic electronic health record environment, and subsequently led the full clinical pilot at Eskenazi.
The result is a body of evidence that traces a single tool from conception through real-world deployment, generated by a research team with continuous institutional access to the data, the infrastructure and the clinical setting.
The value of starting at the build stage
Prior validation studies established that AFib risk models could perform well statistically, but they left several practical uncertainties unresolved. Would clinicians engage with risk information presented during routine visits? Could screening be integrated without disrupting workflow or increasing burden? Would model-enabled identification translate into diagnostic evaluation or treatment decisions in real care settings?
Answering these questions required a partner who understood the model’s architecture and data provenance from the start — not one brought in after to test whether someone else’s tool would survive contact with a real clinical environment. Regenstrief’s continuity across the full project made it possible to design the clinical pilot in a way that was directly informed by what the earlier development and validation work had revealed.
The Answer
The Clinical Pilot
The clinical evaluation embedded UNAFIED into real-world cardiology workflows at Eskenazi Health. Rather than testing the model in isolation, the collaboration focused on the surrounding process:
- Embedding non-interruptive clinical decision support within Epic electronic health record workflows to surface elevated AFib risk
- Pairing model outputs with clinician-led screening using FDA-approved single-lead ECG devices
- Allowing diagnosis and management decisions to follow usual clinical practice
The study examined multiple dimensions of real-world use:
- Identification of patients flagged as elevated AFib risk by the model
- Uptake of screening during routine cardiology visits
- New AFib or atrial flutter diagnoses documented during the study period
- Clinician perceptions of workflow fit, usability, and clinical value
The collaboration demonstrated that an EHR-embedded AFib risk model can be integrated into cardiology workflows without disrupting care delivery. New AFib or atrial flutter diagnoses were documented among patients identified through the workflow, and some patients initiated anticoagulant therapy consistent with guideline-based management.
A follow-up outcomes study examining the cohort identified during proof-of-concept implementation is currently under review.
The analysis was designed to distinguish model-enabled identification from downstream clinical decisions – recognizing that diagnoses and treatment reflect multiple inputs beyond any single tool.
The clinical evaluation embedded UNAFIED, a validated EHR-based model estimating patients’ two-year risk of developing atrial fibrillation, into real-world cardiology workflows at Eskenazi Health. Rather than testing the model in isolation, the collaboration focused on the surrounding process:
- Embedding non-interruptive clinical decision support within Epic electronic health record workflows to surface elevated AF risk
- Pairing model outputs with clinician-led screening using FDA-approved single-lead ECG devices
- Allowing diagnosis and management decisions to follow usual clinical practice
The study examined multiple dimensions of real-world use:
- Identification of patients flagged as elevated AF risk by the model
- Uptake of screening during routine cardiology visits
- New AF or atrial flutter diagnoses documented during the study period
- Clinician perceptions of workflow fit, usability, and clinical value
The collaboration demonstrated that an EHR-embedded AF risk model can be integrated into cardiology workflows without disrupting care delivery. New AF or atrial flutter diagnoses were documented among patients identified through the workflow, and some patients initiated anticoagulant therapy consistent with guideline-based management.
A follow-up outcomes study examining the cohort identified during proof-of-concept implementation is currently under review.
Importantly, the analysis was designed to distinguish model-enabled identification from downstream clinical decisions—recognizing that diagnoses and treatment reflect multiple inputs beyond any single tool.
The relevance for life-science partners
For life-science organizations evaluating where to invest in real-world evidence generation, this collaboration illustrates what becomes possible when a research partner can operate across the full development lifecycle:
- Model development grounded in a longitudinal, multi-institution data asset that reflects diverse patient populations
- Tested in INPC data and subsequently validated in national EHR datasets
- Implementation science expertise to move from a working model to a live clinical workflow, with the study design to evaluate what changes in practice
- A continuous chain of evidence from initial development through outcomes follow-up with the same institutional team
Regenstrief’s role in this project was not to stress-test a finished product in a real-world setting. The organization contributed across the full evidence-generation lifecycle — helping develop and validate the model, integrating it into clinical workflows, and evaluating how it functioned in practice, with Pfizer as a partner throughout the process.
National validation
To evaluate whether the model could perform beyond the regional environment in which it was originally developed, the collaboration externally validated UNAFIED using a large national electronic health record dataset representing more than 100 million patients across diverse healthcare settings.
The analysis examined not only predictive performance, but also bias, transportability and generalizability across patient subpopulations. Findings demonstrated that the model maintained strong predictive performance in a national dataset and informed the development of a more implementation-ready version of the tool designed for broader real-world use.
This phase helped establish whether the approach could scale across different patient populations, healthcare organizations, and electronic health record environments.
Read the Research
Ateya M et al. Validation, bias assessment, and optimization of the UNAFIED 2-year risk prediction model for undiagnosed atrial fibrillation using national electronic health data. Heart Rhythm O2 (2024). DOI: 10.1016/j.hroo.2024.10.004
Grout RW et al. Screening for undiagnosed atrial fibrillation using an electronic health record–based clinical prediction model: clinical pilot implementation initiative. BMC Medical Informatics and Decision Making (2024)
Grout RW et al. Development, validation, and proof-of-concept implementation of a two-year risk prediction model for undiagnosed atrial fibrillation using common electronic health data (UNAFIED). BMC Medical Informatics and Decision Making (2021). DOI: 10.1186/s12911-021-01482-1
Regenstrief’s role in this project was not to stress-test a finished product in a real-world setting. It was to build the product, prove it worked, carry it into practice, and measure what happened — with Pfizer as a partner throughout.