MCH-Guard: Multimodal machine learning framework for risk stratification of cerebral microhemorrhage risk in the Alzheimer's Disease Neuroimaging Initiative.

INTRODUCTION: Efficient cerebral microhemorrhage (MCH) monitoring is critical for anti-amyloid therapy safety due to amyloid-related imaging abnormalities with hemosiderin deposition (ARIA-H) risk. We developed MCH-Guard, a multimodal machine-learning framework, to stratify MCH risk for Alzheimer's Disease Neuroimaging Initiaitive (ADNI) participants (N = 813). METHODS: Nested models integrated clinical history, fluid biomarkers, and imaging to predict MCH presence, incidence, and stability. RESULTS: The comprehensive model detected baseline MCH with high accuracy (area under the curve [AUC] = 0.86). Notably, the minimal model (M1), utilizing only demographics and clinical history, achieved robust performance (AUC = 0.72). Longitudinal models predicted time-to-incidence (R2 = 0.67) and stratified four-year risk. Furthermore, we identified a transient vascular instability phenotype-where MCH status fluctuates-which was strongly predicted by hepatic factors. DISCUSSION: MCH-Guard offers a flexible clinical decision-support tool for optimizing spontaneous MCH and ARIA-H surveillance. The strong performance of the clinical-only model supports equitable risk assessment in resource-limited settings, while the characterization of vascular instability addresses a critical confounder in safety monitoring.