In Clinical Care, AI Still Has to Crack the Value Proposition

MedStar Health SVP and Chief Innovation Officer Bill Sheahan says AI will meet its potential to transform healthcare when it improves clinical outcomes. And that will take some time.

As AI programs reach maturity, so, too, will their value. Early-stage tools that are under the spotlight now for cost will succeed in the long run if they also improve clinical outcomes.

That, says Bill Sheahan, senior vice president and chief innovation officer at MedStar Health and executive director of the MedStar Institute for Innovation, is where AI will be truly transformative. And that's how healthcare executives have to think about the future.

"We believe that the real transformative potential of AI will come from integrated, systemwide adoption," Sheahan, a participant in the HealthLeaders Mastermind program for AI in clinical care, said in a recent e-mail Q&A. "Much like the building of a new hospital within a health system, the long-term impact of AI across our health system will be measured in patient outcomes and margins, not millions."

In a HealthLeaders story last December, Sheahan described how the Maryland-based health system was taking a slow and steady approach to AI, with a particular focus on change management. That process has continued with governance.

"Over the past year, MedStar Health's AI governance has matured from a more exploratory, ad-hoc process into a structured and proactive system," he said. "We launched an AI review process involving experts across the enterprise in innovation, legal, compliance, equity, quality and safety, information security, operations, and beyond. Leaders at MedStar Health are empowered to explore and propose AI tools to address their needs and bring them forward for evaluation."

"The overall volume of new AI products and features being added across all areas of the organization, along with a better understanding of the complexity of integrating AI into clinical care, necessitated different approaches to governance and strategy," he added. "AI that is impacting clinical decision-making or that is patient-facing is typically higher-risk and more complex, requiring more internal expertise from our AI COE (Center of Excellence) than what are typically lower-risk clinical administrative or broader business applications (e.g., coding and billing)."

With that process in place, Sheahan says they're now looking ahead.

Bill Sheahan, senior vice president and chief innovation officer at MedStar Health. Photo courtesy MedStar Health.

"As we further establish our governance processes and opportunities, we increase our focus on strategic imperatives in areas with significant transformational potential that are not yet fully addressed within our current vendor ecosystem, either due to product fit or pricing constraints," he says. "Within these areas, we often buy a solution if offerings in the market are more robust and well-defined, while prioritizing an internal build/partnership model in more nascent areas."

Sheahan and others in the Mastermind program have said it's important to point out that AI isn't exactly new. Traditional machine-learning and predictive modeling have been around for quite some time. The addition of large language models, however, has given a boost to generative AI capabilities.

"In the generative AI space, we are integrating various tools throughout our software stack to support a wide range of application areas, ranging from our safety event tracking system to human resources and informatics," Sheahan says. "Exploration of EHR data is under way, utilizing internal tools to extract and code notes and radiology reports to drive workflows for incidental findings and quality."

"We will also soon roll out an internally-built ‘chat' program in phases across our system," Sheahan adds. "This internal alternative to widely-available tools aims to protect data, improve understanding of usage patterns, and support administrative and clinical staff in searching for system-specific information (e.g., human resources policies). More complex future iterations are expected to integrate patient-level clinical information to allow reasoning over both internal and national clinical guidelines."

Sheahan says the large-scale data warehouses that power large language models are also enhancing the value of traditional predictive modeling.

"Currently, we are implementing a next-generation sepsis algorithm and workflow, with plans to expand to pressure ulcers, fall prediction, and other critical events," he says. "We anticipate that older clinical ‘scores,' such as risk prediction calculators involving only a few simplified variables (e.g., falls, readmission, sepsis, etc.) to inform diagnoses and decision-making, will gradually be replaced with more accurate and fully-automated algorithms. We are also expanding our radiology portfolio to increase the number of findings that tools can detect and use for triaging radiologist review."

At this point in the AI curve, however, ROI is still elusive. There have been some great stories about AI tools that have reduced administrative burdens and workflows and helped both doctors and nurses spend less time on the computer and more time in front of their patients. Sheahan says it will take time for the long-term benefits to show.

"Many of these applications still have limited validation, whether for clinical outcomes or ROI," he says. "As an example, ambient dictation offers the advantage of personal scribes at a fraction of the cost, and providers and patients find it improves the quality of their interactions; however, many health systems are still working to fully quantify and capture the impact needed to secure long-term investment in these products."

"Many of the most promising products are enormously challenging to validate for clinical accuracy or safety as well given current tools, such as large language model products that summarize charts or aid clinicians in reaching diagnoses," Sheahan concludes. "These products otherwise have substantial potential to transform clinical care. Improved frameworks and accepted validation models will be necessary to address safety and outcome questions, leading to greater refinement and broader deployment."

The HealthLeaders Mastermind series is an exclusive series of calls and events with healthcare executives. Please join the community at our LinkedIn page.

To inquire about participating in an upcoming Mastermind series or attending a HealthLeaders Exchange event, email us at exchange@healthleadersmedia.com.