Abstract Number ‐ 18: Potential impact in low and middle‐income countries stroke networks of a deep learning triage tool

Early and accurate identification of large vessel occlusion (LVO) and intracranial hemorrhage (ICH) on initial neuroimaging is essential in a stroke network. A machine learning algorithm (MLA) able to predict LVO or ICH on non‐contrast computed tomography (NCCT) may accelerate workflows.We performed a validation analysis to measure the MLA accuracy among suspected stroke patients transferred to a Comprehensive Stroke Centre (CSC) in Mexico and the possible impact on the workflow in low and middle income countries (LMIC) . From February 2021 to March 2022 consecutive patients with suspected acute stroke who underwent NCCT and computed tomography angiography (CTA) were included. MLA prediction of LVO and ICH was tested against expert physicians readings and clinical follow‐up. We calculated sensitivity, specificity, positive predictive value and negative predictive value. Receiver operating curves were generated for MLA‐LVO, MLA‐ICH and; areas under the curve were calculated. Potential time savings and impact on workflow times were calculated for a scenario in which MLA could analyse initial NCCT at PSC avoiding imaging repetition at CSC. 140 consecutive patients admitted from march 2021 to February 2022 were included in the study, final physicians diagnostics were: 22 ICH (15.7%) and 53 LVO (37.8%) MLA detected 22 ICH (15.7%) and 58 LVO (41.4%).The area under the curve for the identification of ICH with MLA was 0.97 (sensitivity: 94%, specificity: 91%, positive predictive value: 83.3%[MR1][JL2], negative predictive value: 100%). The area under the curve for the identification of LVO with MLA was 0.91 (sensitivity: 100%, specificity: 95.8%, positive predictive value: 85.7%, negative predictive value: 96.4%). Implementation of MLA‐LVO in the network could save CTA acquisition times of 40 (IQR 26) minutes by taking patients directly to the angiosuite for endovascular treatment. In patients with suspected acute stroke, a MLA can quickly and reliably predict ICH and LVO. Such a tool could accelerate the diagnosis, mitigate the contrast imaging dependency and improve the workflow efficiency in stroke networks in LMIC where access to contrast imaging is often limited.