Multicenter evaluation of real-time artificial intelligence assistance for carotid artery stenting
Yuya Sakakura, Ryo Aiura, Kenichi Kono, Yoshikazu Matsuda, Takeshi Fujimoto, Yasunobu Mitsura, Akihiro Sakaguchi, Kazuya Higashizono, Hiroki Nagatsuka, Ikuya Natori, Yoichi MorofujiAbstract
Background
Unintended device movement during carotid artery stenting (CAS) may lead to procedural complications. Our previous preliminary single-center study using a real-time artificial intelligence (AI) assistance system suggested that AI may help detect such movements and offer potential clinical usefulness; however, the small sample size limited the strength of the findings. In this multicenter study, we aimed to validate the accuracy and clinical utility of this real-time AI system in a diverse cohort.
Methods
A total of 53 consecutive CAS procedures with distal filter protection were performed at three institutions under real-time AI monitoring. The AI system continuously analyzed fluoroscopic images and issued notifications when filter devices, guiding catheters, or guidewires moved outside predefined areas or went off-screen. Two types of distal filter protection devices were used. The participating institutions used two angiography systems. The efficacy, safety, and accuracy of the software were evaluated. Subgroup analyses were performed by filter type and operator experience.
Results
On average, there were 13.1 AI notifications per case, including 11.4 true positives and 1.7 false positives. The overall precision and recall were 87% and 98%, respectively. In 32% of true positive notifications, the operator adjusted the inappropriate device position after the notification. No significant differences were observed across filter types or operator experience. No AI-related complications were observed.
Conclusion
This multicenter study demonstrated consistent performance of real-time AI-assisted notifications of device movement during CAS across multiple institutions and operators. The system showed high accuracy and supports its integration into standard neuroendovascular workflows.