Endovascular thrombectomy can reduce disability after ischaemic stroke caused by large vessel occlusion, but access depends on rapid imaging interpretation and efficient transfer to centres that provide the procedure. In England, a large proportion of patients first arrive at primary stroke centres that deliver acute stroke care but do not provide thrombectomy, placing additional pressure on early identification and inter-hospital decision making. National audit data have also shown delays in some elements of specialist stroke care, including admission to a stroke unit within 4 h of hospital arrival for 46.7% of patients. Against this service configuration, AI-based imaging decision support has been deployed within stroke networks to help clinicians identify eligible patients and streamline escalation for time-critical treatment.
AI Decision Support Integrated into Routine Stroke Imaging
The evaluation used data from the Sentinel Stroke National Audit Programme (SSNAP), which records admissions to National Health Service (NHS) hospitals in England for patients aged 16 years and older with a primary diagnosis of stroke. Between Jan 1, 2019, and Dec 31, 2023, SSNAP recorded 452 952 stroke admissions across 107 hospitals. Patient-level analyses focused on 71 017 patients admitted to 26 evaluation hospitals that implemented an AI stroke imaging platform across four regional thrombectomy networks, including 20 primary stroke centres and six comprehensive stroke centres.
The deployed platform, Brainomix 360 Stroke, analysed routine acute imaging used for suspected stroke. This included non-contrast CT and craniocervical CT angiography (CTA), with CT perfusion (CTP) analysed where it was part of local practice. Imaging selection remained governed by each site’s protocol, with CTP and hyperacute MRI predominantly used in comprehensive stroke centres. Workflows were configured so optimised image reconstructions were sent automatically to the AI server alongside standard images used in routine clinical care.
Two periods were used to compare system-level change: a pre-implementation period from Jan 1, 2019 to Feb 29, 2020 and a post-implementation period from Jan 1, 2022 to Feb 28, 2023 by which point all 26 evaluation hospitals had deployed the software. From July 2021, SSNAP added a field recording whether AI support was used for individual patients, enabling comparisons within evaluation hospitals between cases reviewed with and without AI support.
Thrombectomy Rates Increased More Where AI Was Deployed
Across England, thrombectomy rates increased over the 5-year timeframe. The rise was larger in the evaluation hospitals that implemented AI support within their networks. At evaluation sites, the endovascular thrombectomy rate increased from 2.3% in the pre-implementation period to 4.6% in the post-implementation period, a relative increase of 100%. At non-evaluation sites, the rate increased from 1.6% to 2.6%, a relative increase of 62.5%.
Must Read: Enhancing Stroke Prognosis with Deep Learning
Within evaluation hospitals after implementation, AI support was not used for every patient, allowing a direct comparison between groups. In the post-implementation period, 28 670 patients were admitted to evaluation sites, with 15 377 (53.6%) reviewed with AI support and 13 293 (46.4%) reviewed without it. Thrombectomy was delivered more often when AI support was used: 5.9% versus 3.4%. In multivariate analysis, AI usage was associated with higher odds of receiving thrombectomy (odds ratio 1.57). Exploratory analyses indicated that the association was stronger in primary stroke centres than in comprehensive stroke centres.
The evaluation also noted that practice and eligibility criteria evolved during the broader timeframe, including expansion of thrombectomy time windows, while non-evaluation hospitals increasingly adopted AI solutions independently. These factors were considered relevant context when interpreting the national rise in thrombectomy alongside the larger increase observed at evaluation sites.
Transfer Efficiency Improved and Outcomes Shifted Modestly
Transfer from a primary stroke centre to a comprehensive stroke centre is a key operational constraint when thrombectomy capacity is concentrated in specialised sites. Among 747 patients transferred for thrombectomy from evaluation hospitals after implementation, door-in door-out time was shorter when AI support was used. Median door-in door-out time was 128 min (IQR 96–185) with AI support compared with 192 min (119–308) without it. Multivariate modelling identified AI usage as a significant predictor of shorter door-in door-out time.
Use of intravenous thrombolysis also differed by AI usage within evaluation hospitals after implementation. Thrombolysis was given to 17.7% of patients reviewed with AI support compared with 9.0% of patients reviewed without AI support. Over the full timeframe, thrombolysis rates were higher at evaluation than non-evaluation sites, while both showed a gradual decline across the 5 years.
Disability at discharge was assessed using the modified Rankin Scale (mRS), with good outcome defined as mRS 0–2. Good outcome occurred in 47.1% of patients reviewed with AI support and 45.6% reviewed without it, with AI usage associated with an odds ratio of 1.16 for good outcome. In-hospital mortality, defined as mRS 6, was 10.5% in both groups.
Across SSNAP data in England, structured implementation of AI imaging decision support in stroke networks was associated with higher thrombectomy rates, shorter transfer times for patients moved to comprehensive centres for thrombectomy and a modest shift towards better functional status at discharge, while in-hospital mortality was unchanged. The results align with the operational objective of improving identification and referral for time-critical reperfusion treatment, particularly in primary stroke centres where access to specialist interpretation and interventional capability can be more limited. For healthcare decision-makers, the findings reinforce the importance of embedding automated imaging support into routine pathways and ensuring processes enable consistent use alongside efficient transfer coordination.
Source: The Lancet Digital Health
Image Credit: iStock
