BAIVECTOR - stock.adobe.com
How Inova modernized its data architecture to prepare for AI apps
Inova went through a data modernization effort to prepare the health system's data foundation to support the rollout of capabilities such as AI agents and ambient scribes.
Inova, a large nonprofit health system in Northern Virginia and Washington, D.C., that serves more than 4 million patient visits each year across its hospitals, primary care practices and emergency centers, has evaluated more than 300 digital health applications for AI and deployed more than 70 different iterations of AI apps.
To support its use of AI, the health system recently worked with data integration company Fivetran to undergo a data modernization effort. Inova accelerated a four-year data modernization road map to six months to become AI-ready and move implementation from the experimental phase to live production.
To operationalize AI, the health system had to tie together several tech platforms, including Fivetran and DBT, which completed a merger on June 1. Overall, the health system manages more than 1,000 technology contracts. Other partners include Astronomer, Microsoft Fabric and Databricks, which supports the health system's advanced analytics and AI workloads.
The health system's software platforms comprise clinical, imaging, laboratory, financial and Software-as-a-Service solutions. Each of these platforms has its own data models, application programming interfaces and limitations, according to a Fivetran case study.
Inova also developed its own digital AI distribution portal with an Inova-branded look. It plans to provide Inova staff with access to AI agents from other vendor platforms on the health system's portal.
Developing an AI application strategy at Inova
Jon McManus, chief data and AI officer at Inova, said the health system is not trying to compete with third-party app developers when rolling out software throughout the organization; it's just incorporating some of these AI capabilities.
"By centralizing and standardizing data movements with Fivetran, Inova has created an open foundation that supports AI agents, copilots and intelligent applications built on trusted, governed data," the case study stated.
Inova's third-party apps include one that looks for microclotting on chest CT scans post-stroke using an AI algorithm. Additionally, Inova uses the latest models from OpenAI and Anthropic and has developed smaller models or features on top of those models, McManus said.
"Sometimes we want to control the UI [user interface], the look and feel," he explained.
Further, the health system teaches its workers how to be AI-literate and acclimates them to having several AI capabilities as part of their workflow. McManus compares navigating these AI features to using a wayfinding app in a complicated theme park like Disney.
"You need a map; you need to know where the bathrooms are; you need to know there's an app to get reservations on rides, or what the context is," McManus said. "In healthcare, we have to think in a similar way. How do we give people the wayfinding to be AI-literate in the workflows that we're serving these AI capabilities?"
He added that it's hard for healthcare professionals to become AI literate if they must navigate "100 different expressions of AI."
On the customer-facing side, more than 1,000 providers at Inova, primarily in outpatient or emergency settings, use Abridge as an ambient AI scribe, according to McManus. The health system is also piloting AI ambient scribes in some specialty inpatient areas and using agentic AI to refer patients to providers, like dermatologists, who may not be in the Inova network, McManus said.
These AI voice agents send data back to Inova, so that Inova clinicians can help out community partners if patients develop medical complications, McManus said.
Another important AI feature at Inova prompts providers to ask patients questions that help reduce paperwork and authorization burdens. Inova has additional apps under development with similar capabilities. For example, Inova is working on a pilot project with health insurer Elevance Health and Abridge to reduce administrative burden in the precharting workflow, according to McManus.
The project involves Inova sharing the names of Elevance members with upcoming appointments with its providers. It then sends data on these members to Abridge so that providers know what to cover in exams while using the ambient scribes.
"Ensuring that additional information is captured in the note can allow expedited approvals from the payer on services ordered for their care," McManus said.
Inova is also developing AI-driven capabilities that operate in the background and are not patient-facing. For example, it is creating AI tools that search provider schedules for open slots and deliver the data to the health system's access teams, so they can book patients "proactively," McManus shared.
However, as Inova built its IT architecture to support these AI tools, it needed to implement guardrails.
Avoiding vendor lock-in and setting guardrails
As Inova evaluated data modernization partners to become AI-ready, the health system wanted to avoid being locked into vendor contracts, McManus explained.
"We pick our partners carefully," McManus said. "We try and ensure that the underlying techniques and technologies they serve don't make us too single-threaded."
He explained that having a diversity of products in a tech stack rather than working with a single vendor or model developer is important so the health system has flexibility to maneuver and make changes to the IT infrastructure while still protecting the organization's core business logic, which refers to the proprietary rules and algorithms that act like the "brain" of a software program.
Inova turned to DBT Labs to create a shared data infrastructure layer to support data quality and governance. This layer has allowed the health system to improve trust in data across the organization.
McManus further noted that Fivetran has allowed Inova to scale its data infrastructure quickly. The health system standardized data movement with a single, governed ingestion model, which is a way to collect and move data into storage. Inova used Fivetran's Connector SDK to extend the ingestion model throughout the health system.
However, when working with several tech partners, healthcare organizations need consistent goals and architectural designs, he emphasized. Not only that, but healthcare organizations must also create guardrails for AI apps to protect patient privacy.
For example, at Inova, when patients enter the exam room, providers have a conversation with them about the visit being recorded using AI and assure them that the data will not be mishandled, McManus noted.
He further explained that investing in the right vendors to create those guardrails protects not only patients but also the health system's operations. The guardrails allow AI to be adopted responsibly at scale and in accordance with the health system's governance and policy "mechanics," he said.
Ensuring that AI apps remain compliant and adhere to the health system's policies is a key focus, McManus stressed. That means the health system holds every application vendor to the same validation bar.
"Every pipeline, every agent, every model is judged against whether it makes care safer, more equitable, or more accessible," he said. "Governance doesn't have to be an obstacle to AI, but it's really that precondition to be successful. And so we've deeply woven those frameworks together so that we're implementing solutions, and we're governing in a way that maintains reliability, performance and safety to the standards of a regulated health system as carefully as we can."
Brian T. Horowitz started covering health IT news in 2010 and the tech beat overall in 1996.
