In a fictional case, a 74-year-old woman arrives at the emergency department with weakness on the right side of her body and difficulty speaking, which began an hour and a half ago. The non-contrast CT shows no bleeding, and CT angiography is performed next. Before the radiologist opens the images, the team's phone buzzes: the hospital's artificial intelligence (AI) software did not find a large vessel occlusion. The neurologist on call looks at the patient, who still cannot name objects, and at the silent alert.

Programs that read CT angiography and alert the team when they suspect a large vessel occlusion are already part of the routine at many stroke centers. They were made to save minutes, and there is good data that they do. They also make mistakes, and the mistake that worries most is silence in the face of a true occlusion.

Key points

• In an accuracy meta-analysis, AI had a pooled sensitivity of 87 percent and specificity of 95 percent for large vessel occlusion on CT angiography (Susanti et al., 2026).
• Performance dropped in occlusions of the M2 segment of the middle cerebral artery and was weak in the posterior circulation, according to the same review.
• In a cluster-randomized trial at four centers, the software reduced the time between arrival and groin puncture by 11.2 minutes (Martinez-Gutierrez et al., 2023).
• The effect on functional independence at 90 days was inconclusive in that trial, and a meta-analysis of before-and-after studies also found no difference in clinical outcomes.
• At an Australian center, experienced human readers had higher sensitivity than the software, and the program's negative result was not enough to rule out occlusion (Slater et al., 2024).

• What AI looks for on the stroke CT
• How often AI is right: what the meta-analyses say
• Where AI goes wrong: the false negative
• AI versus human readers
• The effect on time to thrombectomy
• Across hospital networks: transfer and telemedicine
• How to use the alert without being used by it
• How NEXO helps train this reasoning
• Frequently asked questions
• Take-home messages

What AI looks for on the stroke CT

In ischemic stroke, a large vessel occlusion is the blockage of a proximal artery, such as the intracranial internal carotid, the M1 segment of the middle cerebral artery or the basilar artery. It is the type of stroke in which mechanical thrombectomy has the most to offer, and in which every minute of delay weighs. That is why the question the team asks in front of the CT angiography is simple and urgent: is a large artery closed?

Automatic detection programs try to answer that question before a specialist is available. In general, they receive the images as soon as they come off the scanner and send an alert to the stroke team's phone. Some compare vessel density between the two hemispheres; others use deep learning networks trained on thousands of scans (Ha et al., 2026).

There are also algorithms that look for the occlusion on non-contrast CT, without waiting for CT angiography. The idea is useful in hospitals where CT angiography is slow or unavailable at night, and it is discussed further below.

The alert is a screening tool, not a report

The starting point for reading any study about these programs is that they were designed as a screening and communication tool. A positive alert serves to call the team and prioritize reading. A negative alert, or the absence of an alert, should carry much less weight, because none of these programs has been validated as a test to rule out occlusion. Anyone who treats the absence of an alert as a normal report turns a speed tool into a source of delay.

How this topic differs from neighboring posts

The post about stroke mimics deals with acute deficit that has no vascular cause, such as a seizure, hypoglycemia or migraine with aura. The text about mild head trauma discusses who needs a CT after a blow to the head. Here the question is different: the patient is already in the CT scanner with suspected ischemic stroke, and a program tries to say whether a large artery is occluded.

How often AI is right: what the meta-analyses say

The systematic review and meta-analysis by Susanti and colleagues included 11 articles, with 10,937 patients, on AI applied to CT angiography (Susanti et al., 2026). The pooled sensitivity was 0.87, with a 95 percent confidence interval of 0.76 to 0.93, and specificity was 0.95, with an interval of 0.91 to 0.97 (Susanti et al., 2026). Heterogeneity between studies was very high on several measures, which means the average figure hides different products, populations and definitions of occlusion.

A sensitivity of 0.87 means that, out of every 100 patients with confirmed large vessel occlusion, about 13 would not be flagged by the software, according to the average estimate. In a service that treats many strokes per month, this is not rare. The authors themselves conclude that negative AI results need careful reassessment, with review of the images and attention to the clinical picture.

What changes depending on the site of occlusion

The same review separated performance by arterial segment. In the M1 segment of the middle cerebral artery, sensitivity ranged from 0.87 to 0.93 in the studies that reported this data, and the pooled area under the curve was 0.913 (Susanti et al., 2026). In the M2 segment, more distal and thinner, sensitivity ranged from 0.30 to 0.86 across four studies (Susanti et al., 2026).

The pooled area under the curve for M2 was 0.72, with an interval of 0.38 to 1.06. The authors reported this result as statistically significant, but the interval includes 0.5, a value that corresponds to a test with no ability to discriminate, and their own sensitivity analysis showed that the significance was not strong: it disappeared when two studies were removed. In practice, performance in M2 is uncertain. The authors also described weak performance in type I internal carotid occlusions and in the posterior circulation.

Non-contrast CT as the first signal

Zhou and colleagues did a network meta-analysis of algorithms that look for large vessel occlusion of the anterior circulation on non-contrast CT (Zhou et al., 2026). The ten included studies were all retrospective validation studies, with 3,632 patients. AI based on the image alone had a sensitivity of 0.78, with a credibility interval of 0.68 to 0.86, and a specificity of 0.88.

Expert and non-expert readers had lower sensitivities, of 0.62 and 0.60, with similar specificity, around 0.86. The sensitivity difference in favor of AI over experts was 0.16, with a credibility interval of 0.03 to 0.29 (Zhou et al., 2026). The authors rated the certainty of this evidence as low and state that it does not establish an effect on the flow of care, reperfusion or functional outcome. For them, non-contrast CT read by AI should work as a signal to call the specialist or prioritize CT angiography, and not as a substitute for it nor as a test to rule out occlusion.

Where AI errs: the false negative

The false negative is the most dangerous error of triage software, because it is silent. The false positive is corrected in minutes by whoever opens the images; the false negative can leave CT angiography in the queue while the patient loses brain tissue.

Why the program lets it pass

Ha and colleagues compared, at a Korean center, two commercial programs in 176 consecutive patients with suspected stroke who underwent CT angiography and perfusion (Ha et al., 2026). Large vessel occlusion was confirmed in 53 patients, and both programs had an area under the curve of 0.93. The authors reviewed the errors one by one.

The most common cause of false positive was severe intracranial stenosis, which mimics occlusion. The main cause of false negative was well-developed collateral circulation in distal occlusions, which masks the point where the vessel is interrupted. When the two programs were adjusted to the same sensitivity of 83 percent, one of them had a specificity of 0.96 and the other of 0.89 (Ha et al., 2026). The choice of threshold therefore shifts the balance between false alarms and missed occlusions.

What software approved for M1 does not see

Catapano and colleagues analyzed 559 patients with suspected stroke whose images went through detection software, at a center in the United States (Catapano et al., 2023). According to the authors, the automatic detection of that program was approved only for M1 segment occlusion, and so the other occlusions were excluded from the analysis. For M1, sensitivity was 71 percent and the positive predictive value was 49 percent (Catapano et al., 2023).

The detail of the restricted approval matters a lot in practice. If the program was made only for M1, the absence of an alert says nothing about a carotid, M2 or basilar occlusion. The authors also showed that adding two simple clinical and imaging signs to the alert, gaze deviation and the hyperdense artery sign, increased specificity and positive predictive value for M1.

The posterior circulation

Basilar artery occlusion is rare, severe and hard to recognize, because symptoms can fluctuate or look like something else. Brito and colleagues tested a deep learning algorithm that looks for posterior circulation occlusion on non-contrast CT, in 184 patients from four institutions in the United States and Europe (Brito et al., 2026).

The software had sensitivity of 55.4 percent and specificity of 80.9 percent, against sensitivity of 27.4 percent and specificity of 91.8 percent for the neuroradiologist (Brito et al., 2026). The program found more occlusions, but erred more on negatives, and even so it missed almost half of the occlusions. The study was retrospective, the positive cases were gathered until a target was reached, and some of the authors work for the software company, which the authors themselves declare.

AI against human readers

The comparison with the radiologist depends on who the reader is, on which occlusion is being sought and on the type of exam. On non-contrast CT, as seen above, the AI was more sensitive than specialists; on CT angiography, the picture can reverse.

At an Australian center, the readers were better

Slater and colleagues compared a commercial software with eight readers at a stroke center in Australia, in 500 patients with suspected ischemic stroke (Slater et al., 2024). The readers were one radiologist, five neurologists and two radiology residents, and the reference standard was the diagnosis of a panel of four interventional neuroradiologists.

For large vessel occlusion, the software's sensitivity was 0.62, and that of the readers ranged from 0.65 to 0.96. When the question also included medium vessel occlusions, the software's sensitivity fell to 0.39, while that of the readers stayed between 0.62 and 0.94 (Slater et al., 2024). All readers had higher sensitivity than the software. The program's negative likelihood ratio for large vessel occlusion was 0.41, and the authors concluded that it was not enough to rule out occlusion.

What this means for you

These comparisons teach you to think of the software as one more test, with its own sensitivity, specificity and likelihood ratios. The same logic as the post on pretest probability applies here: in a patient with aphasia, dense hemiparesis and gaze deviation, the probability of large vessel occlusion is high before any imaging. A negative software result, with a negative likelihood ratio of 0.41, lowers that probability, but it may leave it high enough to require immediate reading of the images.

In a mild and fluctuating deficit, the base rate of occlusion is lower, and a positive alert is more likely to be false.

The effect on time to thrombectomy

If the software is not better than a good reader, the reason to use it is time: the alert reaches the whole team's phones at the same time, while the radiologist may still be on another exam and the neurointerventionist outside the hospital.

The Houston randomized trial

Martinez-Gutierrez and colleagues conducted a stepped-wedge cluster randomized clinical trial at four complete stroke centers in the Houston area, United States (Martinez-Gutierrez et al., 2023). In this design, hospitals enter the software period on randomly assigned dates, and each center contributes patients before and after. The software combined automated reading of CT angiography with secure group messages via mobile phone.

The trial included 243 patients treated with thrombectomy who arrived through the emergency department; those transferred from other hospitals were excluded. The median time between arrival and groin puncture before the software was 100 minutes. With the software, this time fell by 11.2 minutes, with a 95 percent confidence interval of 18.22 to 4.2 minutes of reduction (Martinez-Gutierrez et al., 2023). The time between the start of CT and the start of thrombectomy fell by 9.8 minutes.

And what about the patient outcome?

In the exploratory analysis, the proportion of functionally independent patients at 90 days was 32 percent before and 42 percent after the software (Martinez-Gutierrez et al., 2023). After adjustment, the odds ratio was 1.3, with an interval of 0.42 to 4.0. Because the interval includes 1, the result is inconclusive: the trial was not sized to answer whether the minutes gained changed patients' lives.

The authors note other limitations. The four hospitals belonged to the same health system, with the same medical record and the same protocols, and some physicians worked at more than one center. In services with door-to-puncture times very different from 100 minutes, the effect may be different.

The software depends on who uses it

A later analysis of the same trial, with 101 patients, measured how much the teams interacted with the app (Ebirim et al., 2026). For each one-unit increase in the ratio of comments to patients, the time between arrival and puncture fell by 2.6 minutes, with an interval of 0.13 to 5.09 minutes of reduction. The number of active users per patient, on the other hand, had an inconclusive association, with an interval that included zero.

The gain, therefore, depends on the team talking inside the app, not just on installing it.

What the before-and-after studies say

Sarhan and colleagues gathered, in a meta-analysis, 12 studies with 15,595 patients that compared periods before and after the implementation of the same detection software (Sarhan et al., 2025). The time between arrival and puncture decreased, with a standardized mean difference of −0.50, and the time between CT and thrombectomy also decreased, with −0.71. The authors found no statistically significant difference in clinical outcomes such as hemorrhage, mortality and functional independence at 90 days.

Before-and-after studies have a known problem: between one period and another, much changes in the hospital, from protocols to teams. Part of the improvement may come from these changes and not from the software. That is why the randomized trial in Houston carries more weight, even though it is smaller.

In the hospital network: transfer and telemedicine

Many patients first arrive at a hospital without thrombectomy, where the time that matters is the time to decide on transfer.

A prospective study with telemedicine

Olive-Gadea and colleagues conducted a prospective, multicenter, quasi-experimental study at two telemedicine stroke centers that refer patients to a center with thrombectomy (Olive-Gadea et al., 2026). They compared 129 patients transferred before and 136 after the implementation of an algorithm that reads non-contrast CT and CT angiography. The final decision to transfer remained with the remote vascular neurologist.

The primary outcome, the time between arrival at the local hospital and the transfer request, did not change conclusively: median difference of −2.40 minutes, with an interval of −6.16 to 4.48 (Olive-Gadea et al., 2026). The time between imaging and the request fell by 7.16 minutes, and the time between arrival and puncture fell by 31.28 minutes.

The authors' explanation for the larger gain is not the alert itself. The proportion of patients who underwent CT angiography at the local hospital rose from 28 to 81 percent, and repeat scans at the receiving center fell from 79 to 42 percent (Olive-Gadea et al., 2026). Fewer repeat scans meant faster reperfusion.

A database with thousands of alerts

Devlin and colleagues retrospectively analyzed 14,116 stroke alerts handled by telemedicine in 166 hospitals across 17 US states, comparing hospitals with and without an AI platform (Devlin et al., 2024). The median time from arrival to notification of the neurointerventionist was 50.0 minutes in hospitals with AI and 89.5 minutes in hospitals without it.

The same study found an effect in the opposite direction. In hospitals without on-site thrombectomy, the presence of AI was associated with a lower rate of patient acceptance for thrombectomy by the neurointerventionist, an absolute reduction of 10.79 percentage points (Devlin et al., 2024). The authors could not measure imaging times or patient outcomes, and the observational design does not allow separating the effect of the software from differences between hospitals.

How to use the alert without being used by it

Occlusion detection software is a good example of a tool that helps when it is used as a clue and gets in the way when it is treated as a verdict. The post on automation bias describes the mechanism: faced with a system that usually gets it right, people tend to accept the machine's output and seek less information on their own.

Five practical attitudes

• Read the CT angiography, or make sure someone reads it, in every patient with suspected ischemic stroke, with or without an alert.
• Know which arteries your hospital's software was validated and approved for. A missing alert for M2 or basilar may mean nothing.
• Weigh the neurological exam. Aphasia, neglect, gaze deviation and dense hemiparesis raise the probability of occlusion, as the clinical signs added to the software showed in the Catapano study.
• Be more suspicious of the negative alert in the scenarios where the software errs most: distal occlusions, rich collateral circulation and posterior circulation symptoms.
• Use the messaging channel. The time gain in the Houston trial depended on the team's conversation inside the app.

Calibrating trust

Calibrating trust in the machine is similar to calibrating trust in one's own diagnosis: trusting to the right degree for each situation. A positive alert deserves enough trust to call in the team; a negative one, little enough to keep the human read. When imaging and clinical picture do not match, it is worth naming the doubt and seeking the information that resolves it, as in the post on diagnostic uncertainty. In acute stroke, that information is usually the immediate reading of the CT angiography by someone trained.

What other uses of AI teach

Automatic image reading also appears in chest radiography in the search for tuberculosis and in autonomous screening for diabetic retinopathy. In stroke, the difference is the clock: a false negative takes away minutes of brain, not months, which makes the parallel human read even more important.

Source and design	Who was studied	Outcome	Result as reported	Reading
Susanti 2026, review and meta-analysis of accuracy	11 articles, 10,937 patients, CT angiography	Sensitivity and specificity	0.87 (95% CI: 0.76 to 0.93) and 0.95 (95% CI: 0.91 to 0.97)	Good average, high heterogeneity
Susanti 2026, analysis by segment	Studies with M2 data	Sensitivity in M2	0.30 to 0.86	Variable and inconclusive performance
Zhou 2026, network meta-analysis	10 retrospective studies, 3,632 patients, non-contrast CT	Sensitivity of AI and of specialists	0.78 versus 0.62	AI more sensitive, low certainty
Slater 2024, single-center diagnostic study	500 patients in Australia	Sensitivity for large vessel	Software 0.62; readers 0.65 to 0.96	Readers better
Martinez-Gutierrez 2023, cluster-randomized trial	243 patients at four centers	Arrival to puncture	Reduction of 11.2 min (95% CI: 4.2 to 18.22)	Time gain
Martinez-Gutierrez 2023, exploratory analysis	Same sample	Independence at 90 days	OR of 1.3 (95% CI: 0.42 to 4.0)	Inconclusive
Olive-Gadea 2026, quasi-experimental study	265 transferred patients	Arrival to transfer request	−2.40 min (95% CI: −6.16 to 4.48)	Inconclusive

Sources: Susanti et al., 2026; Zhou et al., 2026; Slater et al., 2024; Martinez-Gutierrez et al., 2023; Olive-Gadea et al., 2026. 95% CI: 95 percent confidence interval, or credibility interval in the Zhou study (Zhou et al., 2026); OR: odds ratio; min: minutes.

Back to the case

The patient in the fictional case had aphasia and right hemiparesis, a picture with a high probability of large vessel occlusion before any imaging. The neurologist did not wait for the alert: he opened the CT angiography on his own phone and called the radiologist to read it with him. The two saw an interruption of contrast in an M2 branch of the left middle cerebral artery, with good collateral circulation, exactly the scenario in which studies show the most false negatives.

The neurointervention team was called and discussed the indication for thrombectomy. That night, the software's silence was the clue that demanded the most attention.

How NEXO helps train this reasoning

In NEXO, the case of the day is fictional, with a real clinical presentation and up to six progressive clues, and you propose the diagnosis by searching for the ICD-10 code. Being suspicious of a result that does not match the clinical picture is the same exercise of treating each piece of information as a clue and reviewing the hypothesis, as in training differential diagnosis with clues.

The limits of use of the game are described on the NEXO Economy page. If you research the teaching of clinical reasoning, the page for Researchers explains how to get in touch. Professionals who write or review cases find information on Curator and on Creators, and institutions can learn about the proposal on the partnerships page. Searching for the right ICD-10 code completes this training.

Frequently asked questions

Does AI read CT angiography better than the radiologist?

It depends on the reader and the occlusion. At an Australian center, all human readers, including neurologists and residents, had higher sensitivity than the software for large vessel occlusion (Slater et al., 2024). On non-contrast CT, where the task is harder for people, AI was more sensitive than specialists in a meta-analysis of retrospective studies, with low certainty.

If the software gave no alert, can I wait for the report?

In a patient with clinical suspicion of large vessel occlusion, waiting is risky. The programs were designed for triage, the negative likelihood ratio measured at one center was insufficient to rule out occlusion, and distal and posterior circulation occlusions are the most often missed. Ensure human reading of the images as soon as possible.

Does the software improve patient outcomes?

The data show a gain in time, but the effect on outcome is inconclusive. In the randomized trial in Houston, the odds ratio for functional independence at 90 days had an interval of 0.42 to 4.0 (Martinez-Gutierrez et al., 2023). The meta-analysis of before-and-after studies also found no difference in clinical outcomes.

Take-home messages

• Large vessel occlusion detection software is a triage and communication tool, not a report.
• Average accuracy is good on CT angiography, but it drops in M2 occlusions and is weak in the posterior circulation.
• The false negative is silent: with high clinical suspicion, the absence of an alert does not remove the need for immediate human reading.
• A randomized trial showed about 11 minutes less to puncture, and the gain depended on active use by the team.
• The effect on patients' functional independence remains inconclusive.

References

1. Martinez-Gutierrez JC, Kim Y, Salazar-Marioni S, et al. Automated large vessel occlusion detection software and thrombectomy treatment times: a cluster randomized clinical trial. JAMA Neurol. 2023;80(11):1182-1190. [PubMed](https://pubmed.ncbi.nlm.nih.gov/37721738/).
2. Ebirim EC, Le NM, Samaha JN, et al. Workflow improvements from automated large vessel occlusion detection algorithms are dependent on care team engagement. J Neurointerv Surg. 2026;18(2):385-389. [PubMed](https://pubmed.ncbi.nlm.nih.gov/39870518/).
3. Susanti L, Cuandra KN, Tristan CD, et al. Diagnostic test accuracy of artificial intelligence in large vessel occlusion: a systematic review and meta-analysis. Neurol Res Int. 2026;2026:5709868. [PubMed](https://pubmed.ncbi.nlm.nih.gov/42028301/).
4. Zhou H, Liu L, Fu B, et al. AI decision support for emergency triage of large vessel occlusion using noncontrast computed tomography: systematic review and Bayesian diagnostic test accuracy network meta-analysis. J Med Internet Res. 2026;28:e105068. [PubMed](https://pubmed.ncbi.nlm.nih.gov/42785737/).
5. Sarhan K, Azzam AY, Moawad MHED, et al. Automated emergent large vessel occlusion detection using Viz.ai software and its impact on stroke workflow metrics and patient outcomes in stroke centers: a systematic review and meta-analysis. Transl Stroke Res. 2025;16(6):2258-2271. [PubMed](https://pubmed.ncbi.nlm.nih.gov/40335883/).
6. Ha HJ, Ryu WS, Kim BJ, et al. RAPID CTA versus JLK LVO for large vessel occlusion detection: a pragmatic comparison of performance and common pitfalls. Neuroradiology. 2026;68(4):877-886. [PubMed](https://pubmed.ncbi.nlm.nih.gov/41805858/).
7. Slater LA, Ravintharan N, Goergen S, et al. RapidAI compared with human readers of acute stroke imaging for detection of intracranial vessel occlusion. Stroke Vasc Interv Neurol. 2024;4(2):e001145. [PubMed](https://pubmed.ncbi.nlm.nih.gov/41583584/).
8. Catapano JS, Lee KE, Desai SM, et al. Number needing review: a novel metric to assess triage efficiency of large vessel occlusion detection systems. Stroke Vasc Interv Neurol. 2023;3(4):e000527. [PubMed](https://pubmed.ncbi.nlm.nih.gov/41583669/).
9. Brito A, Cendrero J, Tanzi L, et al. Initial experience with a deep learning algorithm for detecting posterior circulation large vessel occlusion on noncontrast computed tomography. Stroke Vasc Interv Neurol. 2026;6(4):e002250. [PubMed](https://pubmed.ncbi.nlm.nih.gov/42404827/).
10. Olive-Gadea M, Rizzo F, Cocho D, et al. CATalyze-AI: accelerating referral for thrombectomy in acute stroke patients using an AI-based software. A prospective, multicentre, quasi-experimental cohort study. Eur Stroke J. 2026;11(7). [PubMed](https://pubmed.ncbi.nlm.nih.gov/42497289/).
11. Devlin T, Gao L, Collins O, et al. VALIDATE: [...] the Viz.ai mobile stroke care coordination platform to limit delays in LVO stroke diagnosis and endovascular treatment. Front Stroke. 2024;3:1381930. [PubMed](https://pubmed.ncbi.nlm.nih.gov/41542251/).