糖心视频

Since the dawn of human spaceflight, scientists have carefully studied the effects of space and microgravity on astronauts. After decades of observations and examinations, one truth is certain: space is brutal on the human body. Muscles atrophy, bones lose mass, limbs stretch and鈥攎ore unknown鈥攅yesight can degrade in ways not yet fully understood.

To better understand the loss of vision caused by space flight, University of California San Diego researchers used U.S. National Science Foundation (NSF) ACCESS allocations on the Expanse system at the San Diego Supercomputer Center (SDSC) to predict who is most at risk for developing eyesight issues鈥攂efore liftoff.

The team鈥攍ed by researchers at the Shiley Eye Institute and Viterbi Family Department of Ophthalmology in collaboration with the Hal谋c谋o臒lu Data Science Institute (HDSI) at the UC San Diego School of Computing, Information and Data Science鈥攗sed artificial intelligence (AI), trained on high resolution eye scans, to predict individuals at highest risk.

According to , around 29% of crew members who participated in short-duration space flights reported a degradation of distance or near-visual acuity. For crew members in long-duration missions, that number spiked to 60%.

In 2017, scientists first used the name Spaceflight Associated Neuro-ocular Syndrome (SANS) to describe this vision degradation caused by space flights. While symptoms鈥攊ncluding optic disk (beginning of the optic nerve which goes to the brain) swelling, vision shifts and structural distortions鈥攐ften resolve within weeks to months post-flight, in some cases they do not resolve for years after long-duration missions.

"Our models showed promising accuracy, even when trained on limited data," said lead author Alex Huang, M.D., Ph.D., a professor of ophthalmology at UC San Diego School of Medicine and HDSI affiliate faculty. "We're essentially using AI to give doctors a predictive tool for a condition that develops in space, before astronauts even leave Earth."

Huang added that tools such as the one his team developed can support risk management and, in the future, preventative measures prior to launch.

The AI system was trained using optical coherence tomography (OCT) scans鈥攎icroscope-like images of the optic nerve鈥攃ollected before and during space flight. The researchers also used data from head-down tilt bedrest studies on Earth. In this procedure, participants lay in a continuous six-degree head-down tilt, 24-hours a day, mimicking the effects of weightlessness by shifting fluids toward the head.

Predicting the unpredictable鈥攚ith help from a supercomputer

To overcome the challenge of limited astronaut data, the team used deep learning鈥攁 form of AI that mimics how the brain processes images. They broke each eye scan into thousands of slices, creating a much larger dataset for training the models. The researchers also used data augmentation and transfer learning, which help the AI generalize from small samples.

Then, with the help of NSF ACCESS allocations on SDSC's Expanse system, the team trained and tested their models. The best version could predict SANS with up to 82% accuracy, using only preflight scans. Even models trained on Earth-based bedrest data performed well, suggesting that SANS-like changes in these simulations closely mirror changes in actual spaceflight.

"One of the most exciting findings was how similar the AI's attention patterns were across both space and Earth data," said Mark Christopher, Ph.D., a UC San Diego ophthalmology data scientist and co-author of the study. "This strengthens the case for using Earth-based models to study space health鈥攁 promising development towards advancing human spaceflight research."

To better understand what the AI was "seeing," the researchers used class activation maps鈥攙isual heatmaps that highlight areas of interest. The models consistently focused on specific eye layers involved in fluid balance and pressure鈥攍ike the back of the eye's retinal nerve fiber layer and retinal pigment epithelium鈥攇iving scientists new clues about the biology behind SANS.

While the researchers caution that their models are not yet ready for clinical use, they see enormous potential in future versions. These AI tools could one day help NASA personalize astronaut care, guide countermeasure development and even predict the severity of optic changes during future long-term spaceflight missions.

"The results and models from this study are early, but it's a strong foundation," Huang said. "With more data and refinement, this could become an essential part of astronaut health planning."

The study is published in the .