Health systems should direct AI investment toward implementation science, not new tool development, Sanjiv Shah, MD, director of research at Chicago-based Northwestern Medicine Bluhm Cardiovascular Institute, said on Becker’s “Cardiology + Heart Surgery Podcast.”
“We already have an abundance of AI tools that can help us diagnose diseases earlier and determine when a patient needs a certain treatment,” Dr. Shah said.
In heart failure, for example, ECG AI tools and echocardiography platforms can already flag whether a patient needs mechanical circulatory support, transplant evaluation or hospital admission. The problem, he said, is that none of them are being fully used. The core issue is an implementation challenge: Clinicians receive AI-generated results and do not act on them — even when tools run on an opt-out basis.
The fix, he argued, is treating delivery as a research problem. His team runs A/B tests comparing Epic pop-up alerts against nurse coordinator outreach to determine which approach actually drives clinical behavior change.
“These implementation testing approaches and focus groups are where we’re leaning in, and I think most places will get their fastest return there,” he said.
The second high-value investment area, Dr. Shah said, is smarter use of existing data. Healthcare organizations generate enormous volumes of clinical data, and larger systems now have a viable path toward building proprietary AI products.
“Stop thinking of yourselves as healthcare delivery organizations only,” he said. “You are also high-volume, high-quality data centers, and you need to liberate that data and make it AI- and machine learning-ready.”
He noted that clinicians are uniquely positioned to identify what tools are needed and increasingly have the AI-assisted development capacity to build them. Rather than hiring large dedicated AI teams, Dr. Shah recommends deploying a small number of “AI ambassadors,” individuals fluent in both clinical workflows and AI capabilities, to identify pain points across departments and translate them into deployable solutions.
