When a large vessel occlusion (LVO) is missed or caught late, the consequences compound by the minute. That's why independent validation of AI stroke-detection tools matters so much — not as a checkbox, but as the difference between a workflow that catches a treatable occlusion and one that lets it slip through.
A new study out of Good Samaritan Hospital in San Jose, California, gives healthcare organizations exactly that kind of evidence. Published in the American Journal of Neuroradiology (AJNR), the DUEL study ran two AI-powered LVO detection platforms side by side, in the same comprehensive stroke center, across more than 1,500 consecutive stroke alerts over two years. Few published comparisons of this kind reach that scale.
The hospital's neurosciences, stroke services, radiology, and quality teams retrospectively reviewed CTA data from 1,589 consecutive stroke alerts, collected prospectively in parallel over the two-year period. After excluding cases not sent to the software or compromised by poor contrast bolus, metal artifact, or brain hemorrhage, 1,523 cases remained. Of those, 147 — about 1 in 10 — had a confirmed LVO, defined as occlusion or high-grade stenosis of the intracranial ICA or MCA M1 segment.
RapidAI caught 144 of those 147 cases. That's a 98% sensitivity. Viz.ai caught 108, or 73% — a gap that's statistically about as unambiguous as it gets (P<0.0001).
The specificity numbers told a similar story, though the margin was tighter: RapidAI correctly cleared 94% of LVO-negative cases, against 91% for Viz.ai (P=0.004).
There's a third number worth sitting with: how often each platform actually produced a result at all. RapidAI processed over 99% of eligible cases. Viz.ai processed 90%. A missed detection and a failed-to-process case land the clinician in the same place — no alert, no flag, nothing to act on — so that 10% gap isn't just a technical footnote. In raw terms, Viz.ai didn't flag 39 confirmed LVOs, and the study authors point out what that can mean in practice: delayed diagnosis, delayed treatment, in a disease where delay is the whole problem.
Most AI performance claims come from development datasets — curated, controlled, built to make the algorithm look good. DUEL wasn't that. It was 1,589 real stroke alerts, in order, with nothing excluded except the genuinely unreadable scans. Messy bolus timing, atypical presentations, the usual variability of a busy stroke service — all of it stayed in.
That design answers a more useful question than "does this algorithm work in principle." It answers "does this algorithm hold up on a Tuesday night in the ED." A single-center study has its limits — replication elsewhere would only strengthen the case — but a two-year, consecutive, unselected cohort of this size is a meaningfully higher bar than most published comparisons clear.
Trust in clinical AI isn't built by a vendor's spec sheet. It's built by studies like this one — run independently, by a hospital's own clinical and quality teams, on their own patients, published somewhere that requires peer review before anyone sees it.
That's the context worth keeping in mind here: DUEL wasn't commissioned by either vendor. It came out of Good Samaritan's own stroke and radiology services asking a practical question — which tool actually catches what it's supposed to catch — and publishing the answer.
For health systems weighing LVO detection software right now, that's a different category of evidence than a feature comparison. It's a record of what happened, case by case, for two years, in a real stroke center.
Q: Does a single-center study like this apply to other hospitals?
DUEL is one of the largest consecutive, real-world comparisons of LVO detection software published to date, which gives it more weight than smaller or retrospective-only studies. That said, it reflects one hospital's stroke volume, imaging protocols, and patient population over a specific two-year window. Results at other institutions — with different scanners, contrast protocols, or patient demographics — could vary, and independent replication at additional sites would help confirm how broadly these findings generalize.
Q: How should a hospital evaluating LVO detection software use a study like this?
A large, consecutive, real-world comparison like DUEL is a useful data point, but it shouldn't be the only one. Hospitals should look at performance across their own case mix, confirm how a platform performs on the occlusion types and imaging protocols most relevant to their patient population, and weigh workflow factors — like processing speed and integration with existing stroke alert systems — alongside raw sensitivity and specificity. Requesting site-specific validation data or a pilot period before full deployment is a reasonable next step for any institution making this decision.
Q: What's the clinical significance of the difference in processing rates (99% vs. 90%) between the two platforms?
A non-processed case and a false negative land clinicians in the same position operationally — no alert, no flag, nothing prompting escalation — even though the underlying causes differ. Non-processing can stem from image quality thresholds, contrast timing sensitivity, or how a platform handles borderline studies, rather than a misread per se.
Sachdev H, Hudson A, Ong K, Marklein S, Flores M. Detection of Large Vessel Occlusion Using AI: Evaluating Performance of RapidAI LVO vs Viz.ai LVO in 1,589 Consecutive Code Strokes (DUEL). American Journal of Neuroradiology. Published online May 14, 2026. DOI: 10.3174/ajnr.A9418. https://www.ajnr.org/content/early/2026/05/14/ajnr.A9418.long