Vital Surfaces 94 Percent of Missed Incidental Findings Using Amazon Nova and Amazon Bedrock

Vital develops AI-powered patient engagement solutions that help hospitals improve transparency, communication, and care outcomes. Its digital platform is used across emergency departments and inpatient settings to surface plain-language updates and clinical summaries for patients.
To power these capabilities, Vital initially built a generative AI workflow using Large Language Models (LLMs) hosted in the cloud. This setup enabled the team to parse unstructured clinical notes and extract essential details such as follow-up instructions and medications. It also translated medical jargon into plain-language summaries—reducing clinician workload while helping patients better understand their care.

As adoption grew, the team began encountering several challenges with its cloud provider. Running inference at scale—processing nearly a billion tokens daily—became increasingly costly and resource-intensive. The infrastructure also involved managing multiple cloud environments, which added complexity and operational overhead.

In addition, the system offered limited flexibility in model selection and slowed down iteration cycles. “We knew generative AI could help close the loop on critical findings,” says Dr. Nicholas Sterling, chief medical information officer and research at Vital. “But we needed a more scalable and cost-effective way to do it.”

After evaluating several approaches to scale clinical data extraction and patient communication, Vital selected Amazon Bedrock on AWS for its architecture that supports compliance with the Health Insurance Portability and Accountability Act (HIPAA), broad model selection, and cost-effective performance. The platform also provided the security, scalability, and development agility needed to support Vital’s growing portfolio of AI use cases.

A key factor in that decision was Amazon Nova Micro, which stood out for its ability to process clinical notes with high consistency and significantly lower token costs. Compared to previous solutions, Amazon Nova Micro delivered nearly 23 times cost savings, making it feasible to scale LLM-powered capabilities across the patient journey.

Now fully integrated into Vital’s architecture, Amazon Nova Micro ingests and classifies free-text medical notes, organizing them into a structured taxonomy. This allows the team to extract essential details—such as follow-up instructions, medication names, and high-risk findings like precancerous lesions—with a high degree of accuracy.

To ensure every AI-generated output is safe, reliable, and clinically sound, Vital employs a layered quality assurance framework. Each response passes through an initial information quality AI model, a core model validated for safety by physicians, as well as a secondary AI quality model that acts as a “judge”, flagging any inaccuracies, sensitive information, or other potential risks before information reaches the patient. This system supports stringent internal quality checks and has consistently delivered over 99 percent accuracy in QA evaluations.

Migrating to Amazon Bedrock also helped streamline operations, consolidate infrastructure within AWS, and reduce vendor complexity. “The move to Amazon Bedrock made it easier to manage data flow, vendor governance, and security,” says Te Riu Warren, chief technology officer at Vital. “It’s helped us speed up development and gain trust from hospital partners faster.”

By migrating its generative AI platform to Amazon Bedrock, Vital significantly reduced costs, simplified operations, and improved developer velocity. The move made it possible to scale high-volume clinical inference workloads securely and efficiently across multiple use cases.

“We’re recovering the kinds of findings that could save lives,” says Sterling. “Generative AI is helping us close the loop with patients in ways that weren’t possible before.”

With the new setup, Vital supports over 3 million patients annually across the United States. Hospitals using the platform now surface 94 percent of frequently missed incidental findings, most of which were previously overlooked by medical staff, helping ensure that patients receive the follow-up care they need. “Being able to detect and explain a potentially life-threatening issue to a patient before they leave the ER is a breakthrough,” says Sterling. “And we can do it at scale, helping clinicians deliver the highest quality of care.”

The platform processes nearly one billion tokens daily and does so 23x more cost-effectively with Amazon Nova Micro than with alternative latency-optimized LLMs. This efficiency has allowed Vital to expand capacity and onboard new sites rapidly.

Vital’s layered approach to model safety and validation has also helped build trust with hospitals and compliance teams. By clearly demonstrating how outputs are reviewed, and risks are managed, the team has passed stringent privacy and security reviews—supporting broader adoption across health systems.

“We’ve only scratched the surface,” says Sterling. “Generative AI is helping us deliver care that’s more transparent, more proactive, and more human.”