AI agents are quickly becoming part of the clinical conversation. Since the release of large language models (LLMs), researchers have explored their potential for diagnosis, triage, and patient engagement. But as this new study highlights, the success of these tools depends on access to accurate, high-quality, longitudinal real-world data (RWD).1
Without trustworthy RWD, AI for healthcare remains limited to synthetic cases or generic chat models. With it, AI can begin to approximate the dynamic reasoning and context of real physician–patient interactions. In this study, Verily Life Sciences partnered with HealthVerity to source Verified de-identified EHR data from HealthVerity Marketplace that included patient demographics, histories, medications, labs, and clinical encounters to generate realistic patient vignettes.1
Simulating patients with real-world EHR data
The research introduces a patient simulator built from real-world electronic health record (EHR) encounters. Using Verified de-identified EHR data from HealthVerity Marketplace, the team constructed more than 21,000 patient records across a diverse set of conditions. From these, they generated 519 patient vignettes to power conversational triage simulations.
Each simulated patient speaks in everyday language, shares symptoms naturally, and reveals details only when asked, mirroring the unpredictability of real clinical visits. This approach provides a safe, scalable way to train and evaluate AI triage agents without compromising patient privacy.
Figure 1: Outline of creating a patient simulator from EHR data. Adapted from Rashidian et al., 2025.1
Why HealthVerity Marketplace data makes the difference
Most previous attempts to test AI triage relied on fully synthetic patients or narrow low-quality datasets. By contrast, HealthVerity Marketplace enables access to the nation’s largest ecosystem of healthcare and consumer data. In this study, HealthVerity Marketplace supplied a balanced sample of de-identified EHR data spanning multiple therapeutic areas, that ensured the AI was tested against realistic scenarios rather than artificial constructs.
The breadth, precision, and stability of HealthVerity Marketplace data made it possible to:
- Generate vignettes rooted in real encounters rather than fabricated cases
- Support multi-turn conversations that reflect the complexity of actual patient journeys
- Provide structured context such as labs, medications, and prior history for higher accuracy in AI reasoning
This is what Verified data looks like in practice: transparent, accurate, and scalable.
Multi-agent AI built from RWD for patient triage
The research also proposed a multi-agent AI system called AI Triage. Instead of relying on a single black-box model, AI Triage uses a coordinated set of agents to mirror physician workflows. These agents collect symptoms, plan data retrieval, generate differential diagnoses, and provide guideline-based care recommendations.
When tested against the patient Simulator, AI Triage produced accurate triage decisions in the majority of cases, with physicians agreeing that the system’s reasoning was clinically appropriate. Notably, in over 95% of encounters, the most likely diagnosis identified by physicians was within the top three proposed by the AI.
Recently, the FDA proposed new guidance on using real-world data to train AI tools. Read more on the guidelines and regulations in our blog.
Verified data as the foundation of AI in healthcare
This study reinforces a critical truth: AI in healthcare cannot succeed without Verified real-world data. HealthVerity Marketplace, the only source for Verified data in the industry, provides the foundation for creating realistic simulations and safely evaluating AI performance.
As conversational AI continues to move from the lab to the clinic, the ability to ground these models in real-world encounters will be the difference between hype and real impact. With unmatched coverage, precision, and stability, HealthVerity Marketplace is the data backbone for building the next generation of healthcare AI.
Explore HealthVerity Marketplace to see how Verified data can power your AI and clinical innovation.
References
Rashidian S, Li N, Amar J, et al. AI agents for conversational patient triage: preliminary simulation-based evaluation with real-world ehr data. arXiv:250604032 [cs:CL]. Published online June 4, 2025. https://arxiv.org/abs/2506.04032
