Artificial intelligence (AI)-derived software to help clinical decision making in stroke

There may be variation in the care pathway at different centres. For example, different types of CT scan may be used depending on local preference. The most effective method of recruiting readers for the experimental concordance study should be considered. This should ensure representation across acute stroke centres and comprehensive stroke centres, all staff groups assessing CT images in stroke, and ensuring that those recruited are truly representative of the wider groups fulfilling the same role.

Evidence should be generated in such a way that it facilitates ongoing or future assessment of AI software in stroke (for example, supporting repeatability for future updates or other technologies) in line with standard 16 of NICE's Evidence standards framework for digital health technologies.

The SSNAP dataset only collects data for confirmed cases of stroke. So, even if it is modified to collect data relating to the AI software being unable to analyse CT images, this may not be generalisable to negative cases, which may have features on CT imaging that make it more or less likely that the software will fail.

It is important to note that certain outcomes from the implementation of AI software that would influence its accuracy or time taken to reach a decision, cannot be easily measured. For example, readers may have different reliance on AI depending on their experience.

Even if AI software speeds up the isolated task of reading images, this may not reduce the time to thrombolysis or thrombectomy, because of unavoidable delays elsewhere in the stroke pathway.

Baseline use of thrombolysis and thrombectomy varies across the NHS, potentially because of process factors and attitudes to judging suitability for thrombolysis. Qualitative work by Allen et al. (2022) discusses further contributing factors, and a clinical expert noted that access to treatment differs by site, and by time of day. So, AI software is only one factor to influence changes in treatment rates, and other sources of variance should be considered during evidence generation.

CT perfusion imaging may guide referrals, especially between 6 and 24 hours after symptom onset, when salvaging brain tissue is less likely. This is increasingly available at acute stroke centres, but analysis support may be needed from comprehensive centres, which could influence uptake of thrombectomy.