Abstract TP104: Agreement Of Hypoperfusion Intensity Ratio On CTP Derived By Rapid AI And Viz AI

Abstract

Background: For anterior circulation large vessel occlusion acute ischemic stroke patients, Hypoperfusion Intensity Ratio (HIR) derived from automated Computerized Tomography Perfusion (CTP) software has been shown to predict the rate of collateral flow, speed of infarct growth, and clinical outcome. Published data have proposed an HIR threshold derived by RAPID AI (Menlo Park, CA) of either <0.5 or <0.4 as best predicting robust collaterals. The agreement between HIR derived from VIZ AI (San Mateo, CA) compared to RAPID AI is unknown. Methods: We identified 107 cases in which the middle cerebral artery was the most proximal site of occlusion on Computerized Tomography Angiography as confirmed by a neuroradiologist, and for whom a CTP performed on RAPID AI was diagnostic (e.g. lacking artifact from head motion, a failed bolus injection, poor scan timing, or low signal-to-noise). Those CTP acquisition data were subsequently processed on VIZ AI. We measured the agreement for HIR dichotomized at both <0.5 and <0.4 between RAPID AI and VIZ AI, along with the overall agreement. Results: For the HIR dichotomized at <0.5, the simple kappa coefficient between HIR derived from VIZ AI compared to RAPID AI was 0.714 (95% CI 0.581-0.846), versus 0.636 (95% CI 0.500-0.772) for the HIR dichotomized at <0.4. The intraclass correlation coefficient (ICC) between HIR derived from VIZ AI compared to RAPID AI was 0.78 (95% CI 0.694-0.844). Conclusions: In our sample of 107 cases, the level of agreement for the HIR derived from RAPID and VIZ dichotomized at <0.5 was in the moderate to very good range (Landis & Koch, 1977), versus in the moderate to good range when dichotomized at <0.4. Based on the ICC, agreement was in the moderate to good range. Further investigation of the agreement of HIR derived from RAPID AI and VIZ AI should be conducted before applying HIR thresholds derived from RAPID AI to CTP studies analyzed by VIZ AI to predict favorable collateralization.