Introduction: AI imaging decision support software is recommended by UK and USA stroke guidelines to facilitate identification and transfer of stroke patients eligible for endovascular therapy (EVT) but the impact on thrombectomy delivery is unclear. This prospective observational study evaluated the impact of Brainomix 360 Stroke software in four stroke networks (28 hospitals) in England’s National Health Service (NHS). The primary outcome was percentage of acute stroke patients receiving EVT (the EVT rate); door-in door-out (DIDO) times were assessed as a secondary outcome. Methods: Data were collected prospectively from the Sentinel Stroke National Audit Programme. The impact of Brainomix 360 Stroke software was assessed in two ways: comparison of EVT rates at the 28 evaluation sites and non-evaluation NHS sites before and after implementation (pre-implementation: Jan 2019-Feb 2020; post- : Jan 2022-Feb 2023); comparison of EVT rates and DIDO times at evaluation sites after implementation in patients for whom AI software was used and in those it was not. Multivariate regressions were used to evaluate whether AI use was a predictor of EVT or DIDO time, accounting for other clinical variables (e.g., age, NIHSS, day of week, time of day, time since onset). Results: The dataset included 71,327 patients from 28 evaluation hospitals. Figure 1 shows the change in EVT rates over time in evaluation (blue) and non-evaluation sites (yellow). EVT rate at evaluation sites increased from 2.3% pre-implementation to 4.6% post-implementation (p<.0001). By contrast, EVT rate at non-evaluation sites changed from 1.6% to 2.6% (p<.0001). After implementation, cases that were reviewed with AI software support were more likely to receive EVT (5.9% vs 3.6%; OR=1.6 [95% CI:1.3-1.9, p<.0001). When limiting the analysis to cases admitted to primary stroke centres, the EVT rate was twice as high in cases reviewed with AI software (4.7% vs 2.2%; OR=2.3 [1.8-3.1], p<.0001). DIDO times were shorter for cases reviewed with AI software (median 127 vs. 192 minutes; p<.0001). Conclusions: This prospective study showed that use of AI imaging software was associated with a higher rate of EVT, both by comparing pre- vs. post-implementation data, as well as by comparing cases for whom AI software was versus was not used. Use of AI software was also associated with faster time to transfer. The results support guideline recommendations for AI software to be used in clinical practice.
Read full abstract