AI-based clinical decision support unloved in practice

Mayo Clinic researchers have found that primary care providers welcome the concept of AI-based clinical decision support (CDS) while preferring not to use the technology—at least as configured for their tryout adoption—in day-to-day practice.

In a study slated to be published in the May edition of the International Journal of Medical Informatics, Santiago Romero-Brufau, MD, PhD, and colleagues describe their work anonymously surveying several dozen clinical staff in three primary-care clinics in Southwest Wisconsin.

Participants, whose number included nurses and social workers as well as physicians, completed surveys before and after implementation of an AI-based CDS system.

The CDS’s specific aim was to identify poor glycemic control in patients with diabetes. When the software flagged these patients, it alerted staff and offered recommendations for intervention.

Prior to the program’s installation, some 14 staff members said they felt unfavorably disposed toward the CDS. Only 11 viewed it in a positive light.

But after they’d used the system, a strong majority, 21 staff members, indicated they were favorably disposed. Only three said they didn’t like it.

And yet, despite the high ratings post-implementation, just 14% said they would recommend the AI-based CDS to peers.

Anecdotal feedback offered to the researchers showed the most appreciated aspect of the CDS was that it “promoted team [dialogue] about patient needs.”

The least appreciated aspect was the “inadequacy of the interventions” recommended by the CDS. These recommendations were widely perceived to be “poorly tailored, inappropriate or not useful,” the authors note.

In their discussion, the authors surmise that AI-based CDS tools perceived negatively by staff “may reduce staff excitement about AI technology, and hands-on experience with AI may lead to more realistic expectations about the technology’s capabilities.”

They urge AI developers to carefully distinguish between tasks well-suited for AI and those best left in the hands of humans.  

AI may be poorly suited to assess nuanced aspects of care that involve personal and social factors and are not easily codified within the EHR,” Romero-Brufau et al. comment. “This reinforces that AI is unlikely to replace clinicians anytime in the foreseeable future. Yet, it may be useful as a tool used by trained clinicians, provided that it is carefully implemented in a way that is integrated within local workflows and provides relevant recommendations to augment the abilities of clinical practitioners.”