Stroke AI is an instructive category because its value cannot be assessed from image-interpretation accuracy alone: the whole point of these platforms is compressing the time between a scan being acquired and the right specialist making the right transfer decision, which means notification speed, team communication and workflow integration matter as much as detection sensitivity, and a platform that reads images superbly but integrates poorly into a hospital's transfer pathway may deliver less real clinical value than one with less sophisticated image analysis embedded more effectively.
The comparison, dimension by dimension
Large-vessel-occlusion detection: the core imaging task, identifying the blocked vessel that defines candidacy for mechanical thrombectomy. Perfusion analysis: quantifying salvageable brain tissue, central to selecting patients who will benefit from intervention within extended time windows. ASPECTS scoring: a standardised measure of early ischaemic change on CT, automated scoring reducing interobserver variability in a metric that influences treatment decisions. Intracranial haemorrhage detection: a distinct and equally time-critical task, since haemorrhage fundamentally changes the treatment pathway. Patient transfer and team communication, mobile alerts: the workflow layer that converts an imaging finding into coordinated action, arguably where this category's real differentiation lives, since the underlying imaging algorithms across stroke-AI platforms increasingly perform at broadly comparable levels while integration and communication design vary more. Regulatory indications and NHS evidence-generation status: each platform's specific authorised scope and its position within NHS evaluation frameworks should be checked directly rather than assumed uniform across the category. Effects on door-to-needle and door-to-groin time: the intermediate process metrics that plausibly link platform adoption to improved outcomes, and the most direct evidence available for most deployments. Clinical-outcome evidence: the hardest and least commonly available evidence tier, whether adoption measurably improves functional outcomes at follow-up, distinct from and beyond process-time improvements.
What "NICE permitted while evidence is generated" actually means
Brainomix, RapidAI and Viz.ai are among the stroke technologies NICE has permitted for use in the NHS while further evidence is generated, an Early Value Assessment mechanism explained fully in this cluster's regulatory-literacy article. This is a genuinely important adoption recommendation, NICE judged the technologies promising enough to merit use under structured evidence collection rather than requiring evidence completion before any access, and it is not equivalent to a conclusion that any specific product has already demonstrated improved patient outcomes or established cost-effectiveness. The distinction matters practically: a service adopting one of these platforms under this framework should understand it is participating in continued evidence generation, with defined data-collection expectations, rather than deploying a technology whose outcome benefit is already settled.
The whole chain, not just the algorithm
The most useful way to understand why analysing model sensitivity alone misses much of the clinical value in this category: scan acquired, algorithm analyses the images, alert generated, specialist reviews the alert, transfer decision made and executed, reperfusion treatment delivered, patient outcome follows. Each link in that chain can independently determine whether a technically excellent detection algorithm translates into faster treatment: an accurate alert that a specialist does not see promptly delivers no benefit, a fast alert routed to the wrong team delivers no benefit, and a rapid transfer decision undermined by transport logistics delivers no benefit either. This is why door-to-needle and door-to-groin time improvements, process metrics that capture the whole chain's performance rather than the algorithm's detection accuracy in isolation, are often the most clinically meaningful evidence available for these platforms, more directly connected to patient benefit than any single accuracy figure.
Choosing between the three
The dimension most services should weight most heavily is not raw detection performance, where the three platforms increasingly compete closely, but integration: how cleanly each fits the specific hospital network's existing transfer pathways, on-call communication structures and regional stroke-network geography, since a platform's theoretical capability matters less than whether it demonstrably shortens the actual chain from scan to treatment within a given system's real workflow.
Frequently asked questions
Does NICE permitting use mean these stroke AI tools are proven to save lives?
Not yet in the fully established sense: the Early Value Assessment framework permits use specifically because the evidence base is still developing, and services adopting under this framework should expect ongoing data-collection requirements rather than a settled outcome verdict.
Which single metric best predicts real clinical benefit from stroke AI?
No single metric is sufficient; door-to-needle and door-to-groin time improvements are the most directly outcome-relevant process metrics currently available for most deployments, more informative than detection sensitivity alone.
How should a hospital evaluate which platform fits its network?
By mapping the platform's notification and transfer-coordination features against the hospital network's actual regional stroke pathway and communication structure, since integration quality often determines real-world impact more than algorithmic detection performance.