Performance of Automated RAPID Intracranial Hemorrhage Detection in Real-World Practice: A Single-Institution Experience. Issue 5 (30th September 2022)

• Record Type: Journal Article
• Authors: Eldaya, Rami W.; Kansagra, Akash P.; Zei, Markus; Mason, Evan; Holder, Derek; Heitsch, Laura; Vo, Katie D.; Goyal, Manu S.
• Is Part Of: Journal of computer assisted tomography. Volume 46:Issue 5(2022)
• Page Start: 770
• Page End: 774
• Publication Date: 2022-09-30
• Subjects: RAPID -- artificial intelligence -- AI -- convolutional neuronal network -- ICH -- NCCT -- intracranial hemorrhage -- pragmatic study -- effectiveness study; Tomography -- Periodicals; Tomography -- Periodicals; Tomography; Periodicals; 616.0757
• DOI: 10.1097/RCT.0000000000001335
• Languages: English
• ISSNs: 0363-8715

Abstract

Abstract : Background and Purpose: Intracranial hemorrhage (ICH) is a common finding in patients presenting to the emergency department with acute neurological symptoms. Noncontrast head computed tomography (NCCT) is the primary modality for assessment and detection of ICH in the acute setting. RAPID ICH software aims to automatically detect ICH on NCCT and was previously shown to have high accuracy when applied to a curated test data set. Here, we measured the test performance characteristics of RAPID ICH software in detecting ICH on NCCT performed in patients undergoing emergency stroke evaluation at a tertiary academic comprehensive stroke center. Materials and Methods: This retrospective study assessed consecutive patients over a 6-month period who presented with acute neurological symptoms suspicious for stroke and underwent NCCT with RAPID ICH postprocessing. RAPID ICH detection was compared with the interpretation of a reference standard comprising a board-certified or board-eligible neuroradiologist, or in cases of discrepancy, adjudicated by a consensus panel of 3 neuroradiologists. Accuracy, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) of RAPID ICH for ICH detection were determined. Results: Three hundred seven NCCT scans were included in the study. RAPID ICH correctly identified 34 of 37 cases with ICH and 228 of 270 without ICH. RAPID ICH had a sensitivity of 91.9% (78.1%–98.3%), specificity of 84.4% (79.6%–88.6%), NPV of 98.7% (96.3%–99.6%), PPV of 44.7% (37.6%–52.1%), and overall accuracy of 85.3% (80.9%–89.1%). Conclusions: In a real-world scenario, RAPID ICH software demonstrated high NPV but low PPV for the presence of ICH when evaluating possible stroke patients.