Assessment of Deep Learning-Based Triage Application for Acute Ischemic Stroke on Brain MRI in the Emergency Room

Backgrounds To improve the outcomes of acute ischemic stroke (AIS) patients, a well-organized triage system is essential in the emergency room (ER). This study aimed to assess the detection performance of a deep learning triage research application for AIS that was newly developed based on brain MRI in the ER. Methods This retrospective study consecutively enrolled 831 brain MRIs including mandatory diffusion-weighted image (DWI) performed in the ER of our institution from April to October 2021. MRIs were analyzed with this triage research application as the index test and results were compared with the gold standard of three neuroradiologists. We evaluated sensitivity, specificity, area under the receiver operating characteristic curve (AUROC), area under the precision-recall curve (AUPRC), and maximum F1 score for this application. We also compared changes in its detection performance with and without the addition of optional sequences, which were T1- and T2-weighted images. Results 831 individuals (mean age, 64 years ± 16; 399 men, 432 women) were enrolled and 201 were positive for AIS. The application showed detection performance as follows; sensitivity, 90%; specificity, 89%; AUROC, 0.95 (95% confidence interval, 0.93-0.96); AUPRC, 0.91 (95% confidence interval, 0.86-0.94); and maximum F1 score, 0.86. The addition of optional sequences led to inferior performance compared to the mandatory sequence alone for detecting AIS. Conclusions The triage research application accurately detected AIS in a real-world ER with no additional benefits to its detection performance with optional sequences. Therefore, the triage research application can potentially help clinicians detect AIS in the ER sufficiently/adequately with only mandatory DWI.