Research team conducted multinational validation of AI stroke diagnosis in Korean and US patients, securing the golden time in the emergency room

Dr. Chi Kyung (CK) Kim of the Department of Neurology, Korea University Guro Hospital, and his research team (Dr. Kim Beom-joon of the Department of Neurology, Dr. Jun Sun-woo of the Department of Radiology, Seoul National University Bundang Hospital, and CMO Ryu Wi-sun of JLK) successfully conducted a multinational validation of AI technology that precisely detects large vessel occlusion (LVO) in acute stroke patients using only standard brain CT (non-contrast CT) without contrast agents.

Acute ischemic stroke (cerebral infarction) is a disease caused by blocked blood vessels in the brain, and endovascular thrombectomy is performed to open the blockage, which is crucial to the patient's survival and prognosis. Therefore, CT angiography with the administration of a contrast agent is necessary, but such procedures are often delayed or impossible due to the emergency room environment or hospital system. To this end, there has been a persistent need for technology to rapidly identify large vessel occlusion using only a basic examination method of non-contrast CT.

In order to evaluate the diagnostic performance and clinical utility of JLK's non-contrast CT-based large vessel occlusion identification AI algorithm, the research team analyzed multinational cohort data from a total of 963 patients, including 723 Korean patients and 240 U.S. patients. The results showed that the Area Under the Curve (AUC) representing the diagnostic performance of the AI for large vessel occlusion was 0.963 for the Korean patient group and 0.899 for the U.S. patient group. It also confirmed its usability in different clinical settings by maintaining high diagnostic performance even with patient data from different countries and various CT scanning equipment.

The effect of AI's support was also evident in a cross-validation study with eight medical professionals, including specialists and residents. When medical staff read non-contrast CT scans without AI, the diagnostic accuracy was AUC 0.718. However, with the help of AI, the result improved to 0.852. The sensitivity of accurately identifying patients with large vessel occlusion increased from 46.6% to 63.7%, and the specificity of distinguishing patients without large vessel occlusion also increased from 91.9% to 94.9%.

The research team explained that these results mean that for every 18 non-contrast CT scans read by medical staff, one additional patient with large vessel occlusion that could have been missed with conventional methods can be identified with the help of AI.

Additionally, no clear evidence of automation bias—the tendency of clinicians to uncritically accept AI-generated results—was observed. The research confirmed that AI can serve as a safety net to complement clinicians' judgment during emergency stroke care.

Dr. Chi Kyung (CK) Kim explained, "This is clinically very significant; we can now rapidly and accurately identify severe stroke patients in the time-sensitive emergency room using only the most accessible equipment: non-contrast CT."

Dr. Kim Beom-joon emphasized, "This research is meaningful as it statistically proved that AI can successfully support the medical staff's judgment and can work as a great safety net to compensate for human errors without 'automation bias' that can distort clinical judgment."

Dr. Jun Sun-woo predicted, "With the help of AI, inexperienced medical staff or resident doctors can achieve diagnostic accuracy close to skilled specialists. This will significantly raise the standard level of local hospitals or nighttime emergencies."

This research paper was recently published in the online JNIS (Journal of NeuroInterventional Surgery), a prestigious international journal in the top 10% (Q1) of the neuroimaging and neurosurgery fields.