How often do AI alerts meaningfully impact patient care?

Wednesday, December 2 | 9:30 a.m.-10:30 a.m. | W3-SSIN05-5 | Room N227B

In this scientific presentation, researchers will highlight the real-world impact of AI alerts for flagging of potentially missed radiology findings.

Presenter Charit R. Tippareddy, MD, of UH Radiology Cleveland and colleagues evaluated the real-world clinical impact, potential harm, and operational burden of AI “non-engagement” alerts. These alerts prompt radiologists to re-evaluate a case when they barely interacted with a positive AI result that went undocumented in the report.

The group retrospectively analyzed 154 alerts generated over one year after a commercial AI platform was deployed across a large health system.

Just 19 alerts (12%) resulted in meaningful management changes. Alerts for pulmonary embolism and intracranial hemorrhage were the most likely to change care, although some pulmonary embolism alerts led to treatment complications, including a major bleeding event, according to the researchers.

However, the most common type — rib fracture alerts — had no meaningful clinical impact, and alerts for several other findings likewise didn’t change management. Yet every alert triggered a multi-step workflow managed by radiology informatics staff, making even low-yield alerts disproportionately time- and resource-intensive, the researchers observed.

“AI alert systems should be evaluated based on real-world outcomes — not detection alone — to reduce alert fatigue, avoid unnecessary interventions, and ensure meaningful improvements in patient care,” the researchers wrote.

Find out more at this Wednesday morning session.

Page 1 of 2
Next Page