Automated Intracranial Hemorrhage Detection: Real-World Experience in a Large Comprehensive Stroke Center

Automated intracranial hemorrhage (ICH) detection tools are widespread, yet data are limited regarding their performance in real-world practice.

We retrospectively analyzed noncontrast CT head images of consecutive code stroke patients from January 2022 to February 2023 at a comprehensive stroke center. Patients were included if their indication was stroke, and images were assessed by the automated platform. Radiology reports were considered the gold standard. The primary outcome was the performance of the automated software tool compared with that of board-certified radiologists in ICH detection. The sensitivity, specificity, positive predictive value, negative predictive value, and accuracy were calculated using SAS software.

Of 1,434 code stroke CT scans, 1,402 (98%) were analyzed using the automated detection tool. Thirty-two studies were considered nondiagnostic because of severe motion degradation, and these were excluded. The mean patient age was 67 ± 16 years, with 51% of patients being women and 65% of patients being White. The software tool accurately identified 105 of 129 ICH cases (81%) and 1,255 of the 1,273 non-ICH cases (99%). The sensitivity, specificity, positive predictive value, and negative predictive value with 95% CIs were 81% (74%-88%), 99% (98%-99%), 85% (78%-91%), and 98% (97%-99%), respectively. Automated software sensitivity was highest for intra-axial hemorrhage (IAH) at 94%. Extra-axial hemorrhage (EAH) had a sensitivity of just 43%. Sensitivity of mixed IAH and EAH was 89%, representing a significant increase from isolated EAH.

Automated ICH detection software demonstrates high accuracy in detecting ICH in real-world practice. Sensitivity for IAH is particularly high, with detection of EAH reaching acceptable parameters when co-presenting with IAH but lacking in isolation.