Evaluating and Mitigating Carbon Dioxide Equivalent Emissions in Stroke Management: A Modeling Study

Artificial intelligence (AI) can improve stroke imaging workflows, but its computational carbon footprint remains poorly quantified. We estimated carbon dioxide equivalent (CO₂eq) emissions from AI use in US stroke management and evaluated carbon-aware mitigation.

We combined TriNetX neuroimaging utilization data with Global Burden of Disease stroke burden estimates for 2018 to 2019. Imaging utilization rates were aligned by state, year, stroke type, sex, and age group. We modeled 2 AI scenarios: Minimal AI, defined as one 2-dimensional classification model and one 2-dimensional segmentation model; and Ideal AI, defined as two 2-dimensional classification models, two 3-dimensional segmentation models, and one 3-dimensional classification model. Model energy use was converted to CO₂eq using state-level grid carbon intensity from Electricity Maps.

Ideal AI generated 16 375.09 metric tons CO₂eq/year (95% CI, 16 340.83-16 409.36), compared with 7692.70 metric tons CO₂eq/year (95% CI, 7676.08-7709.32) for Minimal AI. Computed tomography angiography and magnetic resonance imaging were the largest modality contributors. Per-capita emissions varied substantially by state, reflecting local grid composition. Relocating AI-related processing to cleaner-energy states reduced modeled Ideal AI emissions to 7408.68 metric tons CO₂eq/year (95% CI, 7386.99-7430.38), a 54.75% reduction.

AI-assisted stroke imaging has a measurable carbon footprint that depends on model intensity, imaging modality, and grid carbon intensity. Carbon-aware routing of computation offers an immediate mitigation strategy while preserving clinical AI use.