Assessing Real-world Artificial Intelligence Use Among United States Stroke Centers
Nicholas Buonafede1, Diandra Adu-Kyei2, Jaan Nandwani2, Mandip Dhamoon2, J Mocco2, Nathalie Jette3, Laura Stein2
1Mamaroneck High School, 2Mount Sinai School of Medicine, 3University of Calgary
Objective:
We assessed real-world use of advanced imaging in acute stroke protocols and characteristics of US stroke centers that do and do not use artificial intelligence (AI).
Background:
Advanced imaging with AI and perfusion technology can automate assessment of large vessel occlusion and tissue damage in acute stroke. Extent of real-world use is not well known.
Design/Methods:
We utilized an ongoing electronic survey of United States hospitals that treat with thrombolysis and/or thrombectomy (EVT). Hospitals with publicly available electronic contact information were contacted, and respondents included stroke directors or coordinators. We performed retrospective cross-sectional analysis of 20% (n=141) of responding stroke centers at the time of interim analysis. We analyzed questions assessing imaging utilization, center certification and resources, self-reported stroke treatment times, and stroke treatment volumes. We used descriptive analysis and chi-square tests of independence to compare centers that utilize AI to centers that do not.
Results:
Among 141 stroke centers, 84.4% (n=119) utilized AI and 15.6% (n=22) did not. Of the AI centers, 79.8% (n=95) performed EVT in-house and 18.5% (n=22) transferred patients to outside hospitals for EVT. Centers that utilized AI were more commonly comprehensive or thrombectomy-capable centers (72.27% vs 18.18% p=<0.001), served populations of >50,000 (77.31% vs 45.45%, p=0.005), and treated >50 stroke patients annually with thrombolysis (68.91% vs 31.82%, p=0.001) and EVT (56.30% vs 18.18%, p=<0.001). Additionally, AI centers more commonly used CT perfusion in their initial acute stroke imaging protocol (32.37% vs 4.55% p=0.03). There was no difference in self-report median door in-door out times between AI and non-AI centers (p=0.98).
Conclusions:
These limited data suggest that AI use is widespread among certified high-volume stroke centers that perform EVT. Efforts should be made to help smaller and lower resourced centers obtain AI to aid in large vessel occlusion detection.