To really make AI useful, health systems should rethink how they use data, researchers say

© LALAKA - stock.adobe.com

Despite widespread use of electronic health records (EHRs), most health systems still struggle to turn patient data into improved care. A new perspective published in npj Health Systems urges academic medical centers to build what’s known as learning health systems — dynamic infrastructures where data from daily care informs faster, smarter decisions.

“Practitioners will be able to more easily and quickly see what’s working and what’s not; and what’s driving up unnecessary costs,” said Peter Steel, M.D., lead author and associate professor of clinical emergency medicine at Weill Cornell Medicine.

The paper, co-authored by researchers at UC San Diego and Weill Cornell Medicine, argues that artificial intelligence (AI) and quality improvement can’t deliver on their promise without the right data infrastructure beneath them. That starts with connecting siloed systems and aligning operations, research and education.

What is a learning health system?

In a learning health system, data from routine care is continuously collected, analyzed and used to guide changes in care delivery. Instead of waiting years for randomized controlled trials or new guidelines, health systems can iterate in near real time. This model allows for faster adoption of best practices, more responsive care pathways and improved outcomes.

But progress has been slow. EHRs were originally designed for documentation and billing, not for dynamic learning. That’s a problem as AI enters the picture.

“AI can only fulfill its promise if it’s built on a foundation of learning infrastructure,” Steel said.

Primary care can’t afford to be left out

For primary care physicians — especially those in academic or value-based settings — this approach could bring long-overdue support. AI tools for care gap detection, risk stratification and preventive planning all rely on structured, high-quality data. A learning health system provides the foundation for those tools to actually work in practice.

Beyond that, better infrastructure could mean less time spent on manual quality reporting or fighting with fragmented systems. Practices would benefit from real-time feedback, integrated care insights and faster evidence adoption.

And in an increasingly competitive landscape, primary care departments that help lead this shift may gain more leverage within larger health systems.

The price tag — and the payoff

Implementing a true learning health system requires investment — in technology, training and cultural change. But the authors argue that those investments are strategic, not optional. Delaying could mean falling behind in clinical innovation, funding opportunities and operational efficiency.

Institutions like UC San Diego Health, Penn Medicine and Stanford are cited as early adopters of LHS models, using integrated informatics and quality improvement pipelines to drive transformation.

Rewriting medical training

The perspective also calls for education reform. Medical students and residents still graduate with little training in data science, systems improvement, or AI fluency. The authors argue that these skills must become core competencies — not electives — if the health system is to evolve.

Programs such as the AHRQ-PCORI Learning Health System Centers of Excellence and Cornell Tech’s health tech curriculum are starting to fill that gap. Clinicians who understand both medicine and informatics will be essential to bridging care and continuous improvement.

A clear directive

“Implementing the learning health system is no longer a theoretical goal, but a strategic imperative,” Steel said.

In other words: if AI is the tool, the learning health system is the engine. For primary care leaders and practice managers, the message is clear — systems that learn will outperform those that don’t. And practices that embrace this model today may shape the future of medicine tomorrow.