Microgravity, space radiation, and other harsh environmental conditions during spaceflight can change the human body and alter human performance. To safeguard the well-being of crews, Mission Control at NASA’s Johnson Space Center in Houston has relied on an Earth-based team of physicians, biomedical engineers, and other health care experts who work around the clock to monitor and address crew health issues in real time.
Still, the risks of adverse health events during spaceflight remain significant. As crews embark on a 240,000-mile return to the Moon and a 140-million-mile average journey to Mars, remote healthcare delivery will become even more challenging. The farther crews venture from Earth, the more constrained healthcare supplies will be. On a mission to Mars, round-trip communication delays of at least 40 minutes—as well as potential communications blackout periods for weeks at a time—will render real-time medical support from Mission Control impossible. In a medical emergency the astronauts will need on-board decision support tools to augment their capacity for critical, time sensitive medical decisions.
To prepare crews for long-duration space exploration, a team of physicians, nurse informaticists, engineers, and software developers spent nearly three years developing a medical trade space tool that estimates the likelihood that medical conditions at varying levels of severity will occur during spaceflight. Using advanced probabilistic risk assessment (PRA), this tool is also used to recommend the type and quantity of medical supplies a crew will need on hand for a safe return to Earth. The reliability of PRA requires collecting clinical and epidemiological evidence of medical risk, countermeasures, and outcomes for medical conditions expected to occur in spaceflight. Thus far, PRA provides guidance on the risks and treatment protocols for 119 potential medical conditions. This guidance is entered into the PRA knowledge base as a “clinical finding form” (CliFF) that summarizes the incidence, diagnosis, treatment, prognosis, and expected mission impact of the condition.
Until now, developing a CliFF for a medical condition has been an arduous and time-consuming process. For each CliFF, a team of medical researchers conducted rapid systematic reviews of PubMed Central, NASA Technical Reports Server, and other large libraries containing relevant space-related and general population medical evidence. Each of these searches often yielded more than a thousand publications that needed to be reviewed, interpreted, and ranked based on scientific merit and clinical relevance. A subset of the literature deemed to be of highest quality was then interpreted and summarized by NASA health scientists and clinicians to generate each CliFF.
Enter an artificial intelligence-based process for preparing the CliFF. Built by NASA’s AI and clinical science experts with the support of AI experts from Microsoft, this innovative workflow completes in just a few hours what would have previously taken a team of 3 to 4 medical researchers up to a month to complete. Not only does the workflow aggregate the most compelling medical and scientific evidence from publication libraries in a fraction of the time, but it also ensures that the information in each CliFF represents the most recent advances in medical science. This workflow of evidence mining, summarization, and CliFF generation was built by using Microsoft’s Azure AI cloud computing service on internal NASA servers.
In developing this workflow, NASA harnessed the power of Azure AI services to access an open-source large language model (LLM). This model served as the foundation for querying publication libraries by way of a natural language interface. Findings deemed most relevant to the search query were stored in Azure for future retrieval, effectively narrowing the original set containing thousands of publications into a more focused set of only a couple hundred. This information filtering process significantly increased the cost efficiency of future searches.
Using Azure AI services and the refined publication subset, natural language queries were then used to generate CliFFs for three medical conditions--appendicitis, pancreatitis and respiratory failure. Four NASA clinicians who reviewed the AI-generated CliFFs found them to be essentially equivalent to those created without AI assistance. Furthermore, AI-generated CliFFs offered more complete reference lists. Speed, comprehensiveness, and accuracy of summarization were shown to be key advantages of deploying AI to generate CliFFs; however, mathematical calculations and the design of prompts that approximate the nuances of human logic continue to be challenge.
“Through our collaboration with experts at NASA, we were able to demonstrate how the latest AI tools can be used to mine highly specialized medical content, but in a fraction of the time. Because of this, clinical teams can spend less time on sifting through data, and more time on keeping astronauts healthy and in top performance in space” stated Jane Yu, M.D., Ph.D., a senior advisor for healthcare and scientific computing at Microsoft.
“We are honored to collaborate with NASA in their mission to advance human space exploration,” said Thomas Osborne, M.D., Chief Medical Officer of Microsoft’s Federal Civilian unit. “We are proud to be helping ensure that current and future astronauts are well-prepared for the challenges of deep space travel. With Microsoft continuing to innovate AI technologies at a rapid pace, this work reflects only the beginning of what we can do to improve healthcare delivery in remote, disconnected environments in space and on Earth. Our work will continue to exemplify how cutting-edge technology can be leveraged to achieve extraordinary goals.”
Currently, the AI tool developed by NASA and Microsoft is in a prototype stage, and it is not currently used to make operational decisions. Still, the technology shows great promise in optimizing operational efficiency, leveraging a state-of-the-art AI tool to perform a labor-intensive activity in a fraction of the time with similar accuracy. The goal now is to continue testing its reliability and fidelity over time across numerous medical conditions.
In the near term, Microsoft hopes to extend this technology to other scenarios that could allow skilled specialists to improve their efficiency in day-to-day tasks while they focus on more important work.
Microsoft