Artificial intelligence-based software to support mechanical thrombectomy transfer decision in low-volume primary...

To assess the potential benefit of artificial intelligence (AI) based imaging software in supporting mechanical thrombectomy (MT) transfer decisions in patients with acute ischemic stroke (AIS) referred from low-volume primary stroke centers (PSCs).

Many MT-eligible patients are initially managed in PSCs, which often lack advanced imaging capabilities, stroke imaging expertise, and efficient interhospital image transfer systems. Artificial intelligence-based tools for automated large vessel occlusion (LVO) detection have shown promising results in improving stroke workflow metrics, yet data from low-volume PSCs remain limited.

This study presents a multicenter, retrospective analysis of 109 AIS patients transferred for anterior circulation LVO MT from five low-volume PSCs in Poland over a 53-month period (≤ 1 MT transfer/center/month). Standard imaging was retrospectively assessed using Brainomix 360 (Brainomix USA Inc., Chicago, USA) to assess early ischemic changes, collateral status, and LVO location. Two blinded vascular neurologists independently simulated transfer decisions based on post-processed imaging. Large vessel occlusion detection sensitivity and potential changes in transfer eligibility were analyzed. The workflow time parameters were compared to the comprehensive stroke center (CSC) cohort with a routine AI-assisted evaluation (n = 69). The maximal expected time benefit from AI implementation was also estimated.

Artificial intelligence-based sensitivity for anterior circulation LVO detection was 83.5% [95% confidence interval (CI) 76.5-90.5], significantly higher for M1 than for internal carotid artery (ICA) occlusions (95.2% vs. 63.9%, p < 0.01). Among included patients, 78.9% (95% CI 70.3-85.5) were simulated as eligible and could potentially benefit from shorter workflow times. This is supported by the significantly shorter computed tomography angiography (CTA) to endovascular treatment (EVT) notification time in the CSC cohort with routine AI-assisted imaging compared with the low-volume PSC (11 vs. 48 min, p < 0.01). The median maximal potential reduction in door-in-door-out (DIDO) time was estimated at 30 min [interquartile range (IQR) 4-45). In contrast, 4.6% (95% CI 2.0-10.3) individuals were reclassified as ineligible due to extensive early ischemic changes and poor collaterals, potentially avoiding futile transfer.

Artificial intelligence-assisted imaging may significantly improve transfer decisions and workflow efficiency in low-volume PSCs, particularly in settings without real-time radiological interpretation. Its broader adoption may strengthen MT eligibility assessment within regional stroke networks.