How interoperability impacts GenAI deployment in healthcare

How data readiness and interoperability gaps can slow GenAI deployment

Pramila Srinivasan, Ph.D., CEO of CharmHealth, a healthcare technology solutions provider, believes a lack of interoperability can potentially decelerate GenAI deployment by preventing AI models from accessing data.

"Without a standard way for large language models [LLMs] to communicate with EHRs and other software, each model would need new APIs and custom code just to access basic data, making deployment far more expensive, slow and difficult to scale," she said.

In addition, as patients receive care across multiple settings, data remains fragmented, which means GenAI models will lack the context needed for care decisions, she suggested. This data fragmentation would also raise clinical safety concerns and delay validation and regulatory readiness for AI models.

Interoperability and partner integrations enable cross-system data synchronization, which allows GenAI models to have full clinical context, according to Srinivasan.

A lack of interoperability in data exchange also means health systems will miss data from external systems on emergency visits, inpatient stays, procedures, medications and diagnostic results, Srinivasan said. This missing data can increase risk and limit clinical accuracy.

Although local medical practices can still benefit from GenAI by using it for ambient clinical documentation, visit summarization and workflow automation, data coherence issues occur without interoperability, which can impact clinical decision-making, Srinivasan suggested.

"Even the best models can't help clinicians if they have to navigate a maze of disconnected systems," she continued. "LLMs and other forms of AI can only act on the information they have access to, and when EHRs and clinical software all speak different 'languages,' every integration requires bespoke solutions and months or years of engineering effort."

Thus, GenAI can only improve healthcare if there's progress on interoperability.

"They really have to work together," Srinivasan said. "LLMs only perform well when they can access information across systems."

Addressing the 'Wild West' of GenAI deployment in healthcare

Siloed environments are impacting interoperability and, as a result, GenAI utilization, according to Rema Padman, Ph.D., trustees professor of management science and healthcare informatics at Carnegie Mellon University's Heinz College of Information Systems and Public Policy.

"To really leverage the full value of GenAI for the organization, you do need to solve the problem of interoperability," Padman said.

However, Padman described the deployment of GenAI systems as the "Wild West."

"Every organization is coming up with its own way of doing things," she said. "There isn't quite a systematic framework that can guide the development, deployment and evaluation of such systems."

Health systems are deploying their own versions of GenAI applications, such as Gemini or ChatGPT. For example, Stanford Health Care in California developed ChatEHR, which pings EHRs through a Fast Healthcare Interoperability Resources (FHIR) API to obtain historical data on patients, and UC Davis Health has developed custom AI models on population health.

Newly launched AI tools, such as Anthropic's Claude for Healthcare, aim to close the interoperability gap by connecting to many different databases, including ICD-10 code sets. However, the bottleneck doesn't just lie in the technology but also in standards and processes for integration and deployment of GenAI, according to Shyam Natarajan, Ph.D., cofounder and CEO of AI cancer-mapping platform Avenda Health.

"I think this is where actually standardization can take place, not necessarily in the technical protocols, but in policy standardization processes, like policy statements," Natarajan said. "I know there are standards organizations that are trying to work on this, but it's still like the Wild West, especially once you get outside of academia."

The National Institute of Standards and Technology (NIST) is funding a cooperative center at Carnegie Mellon's Heinz College, called the AI Measurement Science and Engineering Center (AIMSEC), to develop systematic approaches to evaluate AI deployments in multiple verticals, including healthcare, Padman said. At AIMSEC, Carnegie Mellon is looking to facilitate "end-to-end interoperability" in areas such as radiology, pathology and oncology.

In addition, normalization and mapping agents translate formats like FHIR, Health Level 7 (HL7) and Continuity of Care Documents into more standardized clinical concepts.

"When data is integrated across care settings, GenAI can present providers with a complete view of a patient's treatment and response trajectory," Srinivasan said. "Seeing the full range and timing of relapses and remissions across encounters enables deeper insight into which medications and interventions are effective and which are not -- unlocking a level of clinical reasoning that is not possible with siloed data."

How GenAI will evolve as interoperability efforts mature

Data integration is growing to include many data sources, including outpatient and inpatient treatment centers, laboratories, imaging systems, research databases and clinical trials, Srinivasan noted. This variety of data sources will enable GenAI capabilities to evolve from task automation to deeper clinical reasoning. That includes correlating interventions with outcomes over time.

"This richer data foundation allows models to move beyond isolated encounter-based insights and toward an understanding of disease progression, treatment response and variability across individuals and populations," Srinivasan said.

In addition, APIs and HIEs will bring together data from separate systems, allowing GenAI to understand clinical context. More integrated data will also allow health systems to pursue Gen AI-enabled data forecasting, she shared.

"With integrated clinical and financial data, GenAI can support forecasting of treatment costs, utilization patterns and outcomes for individuals or defined patient groups," she said. "This enables more informed care planning, resource allocation and value-based decision-making."

Not only that, but interoperable data will also allow GenAI to support precision medicine and deliver effective personalized care, Srinivasan added.

With broader interoperability, GenAI could become a "reasoning layer" on top of integrated healthcare data, learning continuously from real-world evidence, she said.

Brian T. Horowitz started covering health IT news in 2010 and the tech beat overall in 1996.