Full AI stroke imaging in the US could emit 16,375 tons of CO2 a year, a model suggests

Emissions depended heavily on where the computing happens. Per-patient emissions in the ideal scenario ranged from 0.18 kg for ischemic stroke in Vermont to 6.66 kg in Wyoming, reflecting differences in state power grids. In the model, moving AI processing to the cleanest quarter of states cut ideal-scenario emissions by 54.75%, to 7,408.68 tons.

The figures are model estimates built from 82,121 TriNetX stroke records and 12,249,750 Global Burden of Disease-estimated cases, not measured emissions from hospitals. They cover only computing for computer-vision models, not the footprint of the scans themselves, and the authors note that wider use of CT angiography could raise the footprint of ischemic stroke care.

Researchers combined imaging use rates from the TriNetX network with Global Burden of Disease stroke estimates, matched by state, year, stroke type, sex and age group, to estimate annual procedure volumes. They then measured the energy use of median-energy models from the MONAI library with the experiment-impact-tracker tool, including data-center overhead, and converted it to emissions using each state's grid carbon intensity from Electricity Maps.

Ischemic stroke accounted for 91.9% of the TriNetX records and 80.6% of the estimated cases, so it drove most of the total: 12,919.84 tons versus 3,455.26 tons for hemorrhagic stroke in the ideal scenario. Differences between men and women were negligible, while age differences were statistically significant.

For the mitigation test, states in the lowest quarter for emissions per megawatt-hour, including California, Connecticut, Idaho, Maine, New York, Oregon, Vermont and Washington, were treated as hosts, and processing tasks were assigned among them at random. Per-patient emissions in the ideal scenario fell from 2.96 kg to 1.34 kg, and in the minimal scenario from 1.39 kg to 0.63 kg.

The authors say the estimate sits between the modest energy needs reported for segmentation workflows and the much larger footprints modeled for digital pathology. It is well below the energy used to train the largest AI models. They suggest smaller or multi-task models, federated learning, routing non-urgent processing to low-carbon regions and carbon-cost displays in electronic health records.