Abstract Number: LBA25 Validating Artificial Intelligence Model to Optimize Radiologist Detection of Intracerebral Hemorrhage

Abstract

Introduction Artificial intelligence (AI) has shown to be able to alert the radiologist to the presence of ischemic stroke secondary to large artery occlusion (LVO) as fast as 1–2 minutes from scan completion hence leading to faster diagnosis and treatment. In addition to acute LVO, AI has become increasingly used for various intracranial pathologies. In particular, accurate and timely detection of intracerebral hemorrhage (ICH) is crucial to provide prompt life‐saving interventions. Therefore, we aimed to validate a new AI application called Viz.ai ICH with the intent to improve diagnostic identification of suspected ICH. Methods We performed a retrospective database analysis of 4,203 consecutive non‐contrast brain CT reports between September 2021 to December 2021 within a single institution. The reports were made by experienced neuroradiologists who reviewed each case for the presence of ICH. Medical students reviewed the neuroradiologists’ reports and identified cases with positive findings for ICH. Each positive case was categorized based on subtype, timing, and size/volume via imaging review by a neuroradiologist. The Viz.ai ICH output was reviewed for positive cases by medical students. This AI model was validated by using descriptive analysis and assessing its diagnostic performance with Viz.ai ICH as the index test compared to the neuroradiologists’ interpretation as the gold standard. Results 387 of 4,203 non‐contrast brain CT reports were positive for ICH according to neuroradiologists. The overall sensitivity of Viz.ai ICH was 68%, specificity was 99%, positive predictive value (PPV) was 90%, and negative predictive value (NPV) was 97%. Subgroup analysis was performed based on hemorrhage subtypes of intraparenchymal, subarachnoid, subdural, and intraventricular. Sensitivities were calculated to be 86%, 57%, 56%, and 42% respectively. Further stratification revealed sensitivity improves with higher acuity and volume/size across all ICH subtypes. Meningioma was found to be a common false‐positive finding (3 of 22, 14%). Table 1 provides a summary of the results. Conclusions Our analysis seems to indicate that AI can accurately detect the presence of ICH particularly for large volume/size ICH.