Machine Learning–Enabled Automated Large Vessel Occlusion Detection Improves Transfer Times at Primary Stroke Centers

Abstract

Background Accelerating door‐in‐door‐out (DIDO) times at primary stroke centers (PSCs) for patients with large vessel occlusion (LVO) acute ischemic stroke transferred for possible endovascular stroke therapy (EVT) is important to optimize outcomes. Here, we assess whether automated LVO detection coupled with secure communication at non‐EVT performing PSCs improves DIDO time and increases the proportion of patients receiving EVT after transfer. Methods From our prospectively collected multicenter registry, we identified patients with LVO acute ischemic stroke that presented to one of 7 PSCs in the Greater Houston area from January 1, 2021, to February 27, 2022. Noncontrast computed tomography and computed tomographic angiography were performed in all patients at the time of presentation, per standard of care. A machine learning (artificial intelligence [AI]) algorithm trained to detect LVO (Viz.AI) from computed tomographic angiography was implemented at all 7 hospitals. The primary outcome of the study was DIDO at the PSCs and was determined using multivariable linear regression adjusted for sex and on/off hours. Secondary outcomes included likelihood of receiving EVT post‐transfer. Results Among 115 patients who met inclusion criteria, 80 were evaluated pre‐AI and 35 post‐AI. The most common occlusion locations were middle cerebral artery (51.3%) and internal carotid artery (25.2%). There were no substantial differences in demographics or presentation characteristics between the 2 groups. Median time from onset to PSC arrival was 117 minutes (interquartile range, 54–521 minutes). In univariable analysis, patients evaluated at the PSCs after AI implementation had a shorter DIDO time (median difference, 77 minutes; P<0.001). In multivariable linear regression, patients evaluated with automated LVO detection AI software were associated with a 106‐minute (95% CI, −165 to −48 minutes) reduction in DIDO time but no difference in likelihood of EVT post‐transfer (odd ratio, 2.13 [95% CI, 0.88–5.13). Conclusion Implementation of a machine learning method for automated LVO detection coupled with secure communication resulted in a substantial decrease in DIDO time at non‐EVT performing PSCs.

Keywords

• artificial intelligence
• CT angiogram
• endovascular therapy
• imaging
• machine learning
• stroke