Artificial intelligence imaging decision support for acute stroke treatment in England: a prospective observational
Kiruba Nagaratnam1, Ain A Neuhaus2, Lauren Fensome3
1Royal Berkshire NHS Foundation Trust, Reading, UK.
Background:
Endovascular thrombectomy is a standard of care for patients with large vessel occlusion stroke. Artificial intelligence (AI) imaging software is increasingly used to support identification and selection of patients with stroke for this treatment. We aimed to evaluate the effect of AI stroke imaging software on endovascular treatment in England.
Methods:
This prospective observational study was undertaken with the use of data from stroke units in England's National Health Service (NHS). Data on all patients aged 16 years and older admitted to an NHS hospital with a primary diagnosis of stroke were collected through the national stroke audit registry (Sentinel Stroke National Audit Programme; SSNAP). Endovascular thrombectomy rates and interhospital transfer times were measured through SSNAP for all 107 NHS hospitals admitting patients with acute stroke in England from Jan 1, 2019, to Dec 31, 2023, before and after the systematic implementation of stroke AI software (Brainomix 360 Stroke) in 26 hospitals (six comprehensive stroke centres and 20 primary stroke centres; evaluation sites). Hospital-level data were collected for all hospitals, and patient-level data were collected at evaluation sites. The primary outcome was the proportion of patients with stroke receiving endovascular thrombectomy. Changes in endovascular treatment rates were compared for patients who were reviewed with the use of AI software for image interpretation versus those who were reviewed without AI software.
Findings:
452 952 patients with stroke were admitted to 107 hospitals in England between Jan 1, 2019, and Dec 31, 2023. Patient-level data were available for 71 017 patients with ischaemic stroke who were admitted to one of the 26 evaluation sites. For evaluation sites, the pre-implementation endovascular thrombectomy rate was 2·3% (376 of 15 969 patients) and the post-implementation rate was 4·6% (751 of 15 428 patients), a relative increase of 100%. For non-evaluation sites, the pre-implementation rate was 1·6% (1431 of 88 712 patients) and the post-implementation rate was 2·6% (2410 of 89 900 patients), a relative increase of 62·5% (odds ratio [OR] for the interaction between site and time period 1·24 [95% CI 1·08-1·43]; p=0·0026). At the patient level, use of AI stroke software was associated with an increased likelihood of endovascular thrombectomy (OR 1·57 [95% CI 1·33-1·86]; p<0·0001) compared with patients for whom AI software was not used.
Interpretation:
Stroke AI imaging software was associated with increased endovascular thrombectomy rates across the English NHS. These results support the routine use of AI imaging software in the management of patients with stroke.
Funding:
AI in Health and Care Award from the Accelerated Access Collaborative within NHS England.
