Brainomix 360 Triage ICH

Brainomix 360 Triage ICH is an AI-driven software that analyzes adult non-contrast CT brain scans to detect hyperdense volumes indicative of intracranial hemorrhage, providing rapid notifications to specialists for workflow prioritization in acute settings. Its methodology relies on machine learning algorithms for image analysis operating in parallel to standard care, intended for triage assistance rather than standalone diagnosis, with specific performance characteristics and usage limitations.

Brainomix 360 Triage ICH is a radiological computer-aided triage and notification software designed to assist clinicians in rapidly identifying suspected intracranial hemorrhage (ICH) in adult patients. The system utilizes an artificial intelligence algorithm to analyze non-contrast CT (NCCT) images of the brain acquired within 24 hours of acute symptom onset or last known well time. Its primary function is to provide a notification-only, parallel workflow tool, sending alerts to neurovascular or neurosurgical specialists about potential ICH cases, thereby facilitating prioritization for image review. The device allows for image preview via a mobile application or a web user interface, emphasizing that these compressed images are for informational purposes only and not for diagnostic use beyond notification; clinicians must review non-compressed images on a diagnostic viewer for actual diagnostic decisions. The software is specifically optimized to detect hyperdense volumes in the parenchyma associated with acute ICH but is not intended to detect isolated subarachnoid hemorrhage and has limitations regarding scanner compatibility (validated on GE and Philips scanners) and specific patient conditions like recent neurosurgery or certain intracranial pathologies. A retrospective study involving 341 NCCT scans was conducted to evaluate the standalone performance of the Triage ICH image analysis algorithm against a ground truth established by experienced neuroradiologists. The study demonstrated an overall sensitivity of 89.22% and a specificity of 91.37% in detecting ICH, meeting pre-specified performance goals. Performance was further stratified by ICH subtype, showing varied sensitivity (e.g., 96.61% for Intraparenchymal Hemorrhage, 35.71% for Subarachnoid Hemorrhage). The time to notification for the Triage ICH device ranged from 50 to 126 seconds, reported as substantially lower than the average time to notification for the standard of care in the predicate device. The system processes DICOM standard images on an off-the-shelf server and integrates into the Brainomix 360 Platform, using machine learning for automated ICH detection but does not alter original images or replace full patient evaluation for diagnosis.

AI Regulatory Concerns

Powered by

Jev reads the full FDA letter and returns calibrated concern probabilities per regulatory category. These are AI estimates, not regulatory advice.

• Q1. How does the algorithm's performance, particularly the reported sensitivity for Subarachnoid Hemorrhage (35.71%) and Subdural Hemorrhage (66.67%), impact its clinical utility for comprehensive ICH triage, especially considering the device is not intended to detect isolated SAH?; What specific features of the machine learning algorithm allow it to differentiate hyperdense volumes of acute intracranial hemorrhage from other hyperdensities, such as calcifications or contrast, if a patient inadvertently had a contrast CT?; Given the device is validated on GE and Philips scanners, what is the anticipated performance variability or potential for failure if used with NCCT images from other manufacturers or with significantly different acquisition protocols (e.g., varying slice thickness beyond the tested ranges)?; How does the pre-processing and registration involving image rotation and resampling for the web UI's 'straightened images' potentially affect the visual interpretation of subtle ICH findings compared to the original DICOM images, even for preview purposes?; What are the quantitative differences in algorithm performance when stratifying by the source of scans (Boston Medical Centre vs. 'Other' referral hospitals), and what might account for these differences beyond patient demographics?