Balancing the promise of health AI with its carbon costs

The health care industry is increasingly relying on artificial intelligence - in responding to patient queries, for example - and a new Cornell study shows how decision-makers can use real-world data to build sustainability into new AI systems.

The group has offered a framework - Sustainably Advancing Health AI (SAHAI) - for optimizing AI-related energy consumption and emissions in health care settings. SAHAI considers greenhouse gas emissions generated from AI-enabled patient messaging, as well as water needed for cooling hardware in data centers, and scenarios that could affect the emissions profile of a major health system using such a tool.

"Our framework encourages health care organizations, and also technology developers, to think about these different levers and figure out how to balance the promise of AI in health care with being not only mindful of the ethical side, but also the environmental footprint of AI," said Dr. Chethan Sarabu '09, director of clinical innovation for the Health Tech Hub at Cornell Tech.

Sarabu and Udit Gupta, assistant professor of electrical and computer engineering at Cornell Tech and Cornell Engineering, are co-authors of "Sustainably Advancing Health AI: A Decision Framework to Mitigate the Energy, Emissions, and Cost of AI Implementation," which published Sept. 12 in NEJM Catalyst.

Dr. Anu Ramachandran, emergency medicine physician and postdoctoral fellow in medical informatics at the Stanford University School of Medicine, is the lead and corresponding author. Other contributors are Shomit Ghose, lecturer at the University of California, Berkeley; and Dr. Vivian Lee, executive fellow at Harvard Business School and senior lecturer at Harvard Medical School.

Health care systems in the U.S. are turning to artificial intelligence as a way of alleviating the stress on an overtaxed workforce. Automated replies to patient queries are a major part of health AI, expected to grow to a $187 billion industry in the next five years.

But as helpful as AI might be in lightening providers' loads, it is also creating a burden on energy infrastructure. And when considered at a nationwide scale, AI technologies could severely impact health systems' energy utilization and sustainability goals.

To illustrate the parameters driving carbon emissions of health AI tools, the researchers took the example of an AI-generated messaging application, based on an implementation at a large academic health system. They considered one year of operation, with 3,000 physicians answering 50 messages per day.

They calculate that a year of running the AI-powered messaging tool would produce around 48,000 kilograms of carbon dioxide (CO2), or approximately 2,300 "tree-years." A tree-year is the amount of CO2 that one tree pulls out of the atmosphere in a year; that figure varies depending on a host of factors, but the researchers' calculation is based on around 21 kg of CO2 per tree per year.

Their modeling was done using a lightweight generative pretrained transformer, or GPT, which uses less computing power than a larger model. The lightweight model can be used to direct patients to a provider, or answer general questions, but for more involved tasks, a larger model would be necessary.

Providers, Sarabu said, must consider myriad factors - including energy usage, and water consumption for cooling - when deciding how and when to deploy AI.

"If you're responding to a patient about routine follow-ups, small differences in model accuracy, such as the difference between 83% accuracy and 85% accuracy, may not be noticeable," Sarabu said. "But if you generate double the amount of emissions with 85% accuracy, that's probably not striking a good balance. Of course, these decisions need to be carefully weighed with the task and patient outcomes being evaluated."

It would be much easier, he said, to consider sustainability before a system is built, as opposed to retrofitting after the fact.

"We're really in the early days of AI being implemented in health care," he said, "and what happens in the next three years or so, that's going to get baked into the system. And so if we make energy conscious decisions right now, we'll have a more efficient system."

Gupta, whose background is in computer architecture and systems, said taking sustainability into account doesn't have to come with trade-offs.

"A data center that is placed on a more renewable grid, or has access to renewable energy - that's going to have a huge impact on the operational emissions of that data center," Gupta said. "Hospitals can make decisions on where they want to run these AI workloads, prioritizing data centers that operate on renewable energy."

The researchers conclude that there is "a window of opportunity" to align the large-scale integration of AI with consideration of the environmental costs.

"Although the impacts of climate change weigh most heavily on vulnerable, lower-resource patients and the health systems that serve them," they wrote, "emissions generation is driven largely by high-income countries and high-resource health systems, which must consider and mitigate their contributions. This requires recognizing the scale and drivers of impact … and investing in strategies to mitigate the footprint of AI in health systems."