CATalyze-AI: accelerating referral for thrombectomy in acute stroke patients using an AI-based software. A prospective, multicentre, quasi-experimental cohort study.
Authors
Affiliations (10)
Affiliations (10)
- Stroke Unit, Neurology Department, Hospital Universitari Vall d'Hebron, Barcelona, Spain.
- Departament de Medicina, Universitat Autònoma de Barcelona, Barcelona, Spain.
- Neurology Department, Hospital General de Granollers, Granollers, Spain.
- Emergency Department, Hospital Universitari de Vic, Vic, Spain.
- Interventional Neuroradiology, Hospital Universitari Vall d'Hebron, Barcelona, Spain.
- Emergency Department, Hospital General de Granollers, Granollers, Spain.
- Radiology Department, Hospital Universitari de Vic, Vic, Spain.
- Neurology Department, Hospital Universitari de Vic, Vic, Spain.
- Grup de Recerca en Ictus, Vall d'Hebron Institut de Recerca, Barcelona, Spain.
- Stroke Unit, Neurology Department, Hospital Clinic, Barcelona, Spain.
Abstract
Timely transfer of patients with suspected LVO remains critical in acute stroke systems. Artificial intelligence (AI)-based imaging tools are increasingly implemented to support triage in non-thrombectomy centres. We assessed whether integrating an AI algorithm within established tele-stroke centres reduces time to transfer decision. We conducted a prospective, multicentre, quasi-experimental study comparing consecutive cohorts in 2 tele-stroke centres referring to a single comprehensive stroke centre, before and after implementation of the Methinks Stroke Suite, an AI algorithm for LVO detection on non-contrast CT (NCCT) and CTA. Consecutive transferred patients with suspected acute ischaemic stroke were included. Remote vascular neurologists retained responsibility for final transfer decisions. Each phase spanned approximately 15 months. The primary outcome was time from arrival at the local centre to emergency medical services activation. Secondary outcomes included workflow intervals, imaging utilisation and algorithm performance. We included 265 patients (136 in the post-implementation cohort; 129 in the pre-implementation cohort). Adjusted median time from arrival to transfer request did not differ (median difference - 2.40 min; 95% CI, -6.16 to 4.48). Post-implementation, time from imaging to transfer request (-7.16 min; 95% CI, -13.02 to -1.85) and arrival to groin puncture (-31.28 min; 95% CI, -60.89 to -13.74) decreased. Computed tomography angiography acquisition at referring centres increased (28%-81%), reducing repeat imaging at the comprehensive centre (79%-42%). Non-contrast CT-based AI prediction yielded a positive predictive value of 66% for endovascular treatment. Artificial intelligence implementation was not associated with a shorter time to transfer decision. Fewer redundant imaging examinations were associated with a shorter time to reperfusion.