Implementation of artificial intelligence stroke imaging software was associated with higher endovascular thrombectomy rates at participating hospitals, according to a large prospective observational study.

Artificial intelligence (AI) imaging support was also associated with shorter interhospital transfer times for patients with acute stroke across England's National Health Service (NHS). Hospitals that implemented AI software experienced a doubling of endovascular thrombectomy rates, and patients whose imaging was reviewed with AI were significantly more likely to receive the procedure than those evaluated without it.

Researchers analyzed data from the Sentinel Stroke National Audit Programme (SSNAP), a mandated national registry capturing all patients aged 16 years and older admitted to NHS hospitals with a primary diagnosis of stroke. The analysis included 107 hospitals between January 1, 2019, and December 31, 2023, comprising 26 evaluation sites where AI imaging software was systematically implemented and 81 non-evaluation sites used for comparison. Patient-level analyses were conducted at evaluation sites after July 2021, when registry fields recorded whether AI supported interpretation of acute brain imaging.

Among nearly 453,000 patients admitted with stroke during the study period, patient-level data were available for 71,017 patients with ischemic stroke treated at evaluation sites. At those sites, the proportion of patients receiving endovascular thrombectomy increased from 2% before implementation to 5% after implementation, representing a relative increase of 100%. In contrast, non-evaluation sites showed an increase from 2% to 3%, a relative increase of 63%. Notably, 69% of non-evaluation sites independently adopted AI software during the study period, which may have contributed to their improvement.

At the individual patient level, AI-supported imaging interpretation was independently associated with greater likelihood of endovascular thrombectomy after adjustment for age, stroke severity, premorbid disability, timing variables, and thrombolysis use. The association was strongest at primary stroke centers, where access to specialist neuroradiology expertise is limited. "For those patients presenting to a primary stroke [center] and transferred to a comprehensive stroke [center], the median door-in door-out time was 64 min shorter when AI software was used than when it was not," stated Kiruba Nagaratnam, MD, of the Royal Berkshire NHS Foundation Trust, Reading, UK, and colleagues.

Secondary analyses showed that median door-in door-out time was 128 minutes when AI was used compared with 192 minutes without its use. AI support was also associated with higher rates of intravenous thrombolysis and a modest increase in good functional outcomes at discharge, defined as a modified Rankin Scale score of 0 to 2. No association was observed between AI use and in-hospital mortality.

The researchers noted several limitations. The observational design limited causal inference, and residual confounding may have persisted despite multivariable adjustment and propensity score analyses. Patients reviewed with AI tended to have more severe strokes but better premorbid function, which may have influenced treatment decisions. Long-term functional outcomes could not be evaluated because of limited 6-month follow-up data, and reasons for nonuse of AI in some cases were not captured.

Full disclosures can be found in the published study.

Source: The Lancet: Digital Health