Artificial intelligence-quantified total hemorrhagic burden on susceptibility-weighted imaging in acute ischemic stroke: a multicenter cohort analysis

Background: Susceptibility-weighted imaging (SWI) reveals three forms of cerebral hemorrhagic burden: cerebral microbleeds (CMB), chronic parenchymal hemorrhage, and cortical superficial siderosis (cSS). Prior work has focused on CMB counts, whereas the other two are rarely quantified at scale; their prognostic value in ischemic stroke is thus uncertain. An artificial intelligence (AI) tool to quantify all three at registry scale was developed and tested to determine whether the resulting total hemorrhagic burden (THB) score stratifies risk.Methods: The tool was applied to the Clinical Research Collaboration for Stroke in Korea registry (16 sites). After restricting to first-ever ischemic stroke without intravenous thrombolysis or thrombectomy and pre-stroke modified Rankin Scale (mRS) ≤1, 6,124 patients were included in the analysis. THB summed the z-transformed logged values of the three components.Results: CMB, parenchymal hemorrhage, and cSS were positive in 45.8%, 17.1%, and 3.4% of patients, respectively, with distinct clinical determinants. THB was associated with 3-month poor outcome (mRS ≥3): adjusted odds ratio (OR) per standard deviation 1.19 (95% CI, 1.11–1.27; P<0.001). The ordinal mRS shift was driven by parenchymal hemorrhage volume (joint-model OR, 1.23 per log unit) and CMB count (OR, 1.19); cSS was directionally consistent but non-significant. Poor-outcome rates rose monotonically across THB tiers, from 18.6% in the lowest to 36.8% in the highest tier.Conclusion: An AI tool feasibly quantified all three SWI hemorrhagic findings at the registry scale. The resulting THB composite stratified patients by 3-month functional outcome with a tier-based gradient, thus supporting its use for risk stratification.