Intracranial hemorrhage (ICH) requires rapid and accurate diagnosis. Artificial intelligence (AI) may support ICH detection on brain CT, particularly where radiological expertise is limited.

To assess the diagnostic performance of AI for ICH detection on brain CT and identify factors associated with performance variability.

Following PRISMA-DTA recommendations, diagnostic accuracy studies providing sufficient data to construct 2×2 tables were systematically identified. Independent validation cohorts were the primary analytical units, with overlapping datasets excluded from primary pooling (Forty-six studies). Sensitivity and specificity were jointly estimated using a Bayesian bivariate random-effects model. Bayesian subgroup and meta-regression analyses examined validation setting, reference standard, model purpose, and target condition.

The Bayesian bivariate random-effects model yielded a pooled sensitivity of 91.3% (95% CrI, 89.0–93.3%) and specificity of 95.4% (95% CrI, 93.7–96.6%). The positive and negative likelihood ratios were 19.69 (95% CrI, 14.41–26.99) and 0.091 (95% CrI, 0.070–0.116), respectively, with a diagnostic odds ratio of 216.7 (95% CrI, 131.8–354.3). The sensitivity analysis restricted to eight low-risk-of-bias studies yielded sensitivity of 90.3% (95% CrI, 81.2–95.2%) and specificity of 93.4% (95% CrI, 86.7–96.4%). In separate meta-regression models, expert-based reference standards were associated with higher sensitivity (β = 0.713; 95% CrI, 0.173–1.249; posterior probability, 99.5%), and clinical/real-world validation with higher specificity (β = 0.602; 95% CrI, 0.001–1.175; posterior probability, 97.5%).

AI showed high diagnostic accuracy for intracranial hemorrhage detection on brain CT, but low-certainty evidence and heterogeneity limit generalizability. Prospective, external, multicenter validation should assess clinical outcomes and workflow impact, including in resource-limited settings.