Multicenter validation of artificial intelligence software predicting large vessel occlusion using noncontrast brain CT

Background To validate JLK-CTL, an artificial intelligence (AI) software developed to predict large vessel occlusion (LVO) using non-contrast CT (NCCT) scans, and to investigate its clinical implications regarding both infarct volume and functional outcomes. Methods Between January-2021 and April-2023, a consecutive series of patients who concurrently underwent CT angiography and NCCT within 24 hours of last- known-well (LKW) were collected. LVO was confirmed through consensus among three experts reviewing CT angiography. Infarct volumes were quantified using diffusion-weighted imaging (DWI) conducted within seven days of the NCCT. The performance of the JLK-CTL was evaluated based on the area under the receiver operating characteristic curve (AUROC), as well as its sensitivity and specificity. The association of JLK-CTL LVO scores with infarct volumes and functional outcomes was assessed using Pearson correlation and logistic regression analyses, respectively. Results Of 1,391 screened patients, 774 (mean age 69.0 ± 13.6 years, 57.6% men) were included. The median time from LKW to NCCT was 3.1 hours (IQR 1.5–7.4), with 24.2% (n=187) presenting LVO. The JLK-CTL demonstrated AUROC of 0.832 (95% CI 0.804–0.858), with a sensitivity of 0.711 (95% CI 0.641–0.775) and a specificity of 0.830 (95% CI 0.797–0.859) at the predefined threshold. Incorporating the National Institute of Health Stroke Scale into the model increased the AUROC to 0.872 (95% CI 0.846–0.894; p<0.001). The LVO scores showed a significant correlation with infarct volumes on follow-up DWI (r=0.53; p<0.001). When JLK-CTL LVO scores were categorized based on observed frequency of LVO, the highest JLK-CTL LVO scores (51-100) group showed an independent association with unfavorable functional outcomes (adjusted odds ratio 9.48; 95% CI 3.98–22.55). Conclusion The performance of the AI software in predicting LVO was validated across multiple centers. This tool has the potential to assist physicians in optimizing stroke management workflows, especially in resource-limited settings.