Stroke AI: One More Catch Per 18 Scans

The scan that says maybe

It is 2 a.m. A patient comes in with slurred speech and a droopy arm. The stroke team orders a CT, but there is no time and sometimes no machine for a full CT angiogram. What they have is a plain, non-contrast CT: fast, cheap, available in almost every hospital that has a scanner at all. The problem is that a plain CT is bad at showing a large vessel occlusion (LVO), the kind of clot that causes the most disabling strokes. Radiologists and neurologists read subtle density changes and hope they are not missing the one case that needed an emergency thrombectomy.

A new multinational study says an AI model can close a real chunk of that gap, using nothing but the scan the hospital already has.

What the study actually found

Researchers led by Dr. Chi Kyung Kim at Korea University Guro Hospital, working with teams at Seoul National University Bundang Hospital and the AI company JLK, tested an LVO-detection model on 963 patients: 723 in Korea, 240 in the United States. Published in the Journal of NeuroInterventional Surgery, the results are specific enough to sit with:

• The AI's own diagnostic performance hit an AUC of 0.963 in the Korean cohort and 0.899 in the U.S. cohort.
• Clinicians reading scans without AI support scored an AUC of 0.718. With the AI's output alongside them, that rose to 0.852.
• Sensitivity, the ability to actually catch a true LVO, went from 46.6% to 63.7% with AI support.
• Specificity moved from 91.9% to 94.9%, meaning it did not just get more trigger-happy.

The headline number is simpler: for every 18 scans a clinician reviewed with AI support, the tool surfaced one additional occlusion that would otherwise have been missed.

Why non-contrast CT is the interesting part

Most AI stroke tools live on top of CT angiography or perfusion imaging, studies that need contrast dye, more time, and more equipment. Plenty of hospitals, especially at night, with a resident on call and a radiologist reading remotely, do not have that pipeline moving fast enough. Dr. Kim's framing is blunt: this works on "the most accessible equipment," the plain CT that almost every stroke pathway already starts with. Dr. Kim Beom-joon at Bundang put it as a safety net, not a replacement for the reading clinician.

That distinction matters more than the AUC numbers. The model is not making the call. It is flagging the scan that a tired resident, three hours into an overnight shift, might read as equivocal and move on from. The study's authors go further, arguing that this kind of support could let less experienced staff approach the accuracy of stroke specialists, which is exactly the gap that shows up on night shifts and in hospitals without a dedicated neuro-ICU team on site.

What it does not settle

This is a retrospective validation and reader study, not a trial of patient outcomes. It tells you the model finds more true occlusions and does not flood the list with false alarms. It does not yet tell you whether that translates into faster thrombectomy, fewer missed treatment windows, or better recovery scores at 90 days. Two cohorts, two countries, and a still-modest 240-patient U.S. arm is a real start, not a finish line.

Why this is the shape of AI that matters on the ward

The doctor at 2 a.m. does not need a model that is right in the abstract. They need one that catches the specific case they are at risk of missing, on the equipment they already have, without adding a workflow step they do not have time for. That is the bar clinical AI has to clear to be useful on a shift, not just in a benchmark. It is also the bar Medows holds its own AI to: verifiable, tied to the actual case in front of the doctor, not a black box promising accuracy in general.

Sources

• Stroke diagnosis improved by AI support, multinational study finds (Medical Xpress)
• AI assisted detection of large vessel occlusion on non-contrast CT: multinational validation and reader study (Journal of NeuroInterventional Surgery, DOI 10.1136/jnis-2026-025339)