18 September 2025
A new study published in American Journal of Ophthalmology has described a deep learning model for predicting which astronauts are more at risk of vision problems in space.
NASA reports that around one in three (29% of) astronauts experience worsening distance or near vision on short-duration flights, while 60% of those on long-duration flights experience the same problems.
These vision changes have been described as Spaceflight Associated Neuro-Ocular Syndrome (SANS).
Researchers from the University of California San Diego used pre- and in-flight optical coherence tomography (OCT) scans to train an artificial intelligence (AI) deep learning model in identifying which astronauts are most at risk of SANS.
To overcome the challenges of working with a relatively small dataset, the researchers broke each OCT scan into thousands of slices and used data augmentation and transfer learning techniques.
The scientists developed a model that could predict SANS with 82% accuracy from pre-flight OCT scans.
Professor of ophthalmology at UC San Diego School of Medicine, Alex Huang, highlighted that the model shows a promising degree of accuracy – even when trained on limited data.
“We’re essentially using AI to give doctors a predictive tool for a condition that develops in space, before astronauts even leave Earth,” he said.
Comments (1)
You must be logged in to join the discussion. Log in
Don Williams18 September 2025
Fascinating application of deep learning to a clinically unique challenge such as SANS. The use of OCT volumetric slicing, data augmentation, and transfer learning to overcome the classic “small-n, high-d” problem in astronaut datasets is particularly elegant. An 82% predictive accuracy from pre-flight scans really speaks to the value of leveraging convolutional architectures in a domain where sample scarcity would normally hinder model generalisation.
I wonder if the team explored ensemble methods or hybrid models (e.g. combining CNN-extracted features with Bayesian inference) to address uncertainty calibration? Given the heterogeneity of SANS progression, questions around model explainability (e.g. Grad-CAM heatmaps or saliency maps) also become critical if clinicians are to trust AI-driven risk stratification.
This feels like a perfect example of Augmented AI in healthcare — where human expertise is enhanced rather than replaced. It would be exciting to see how similar architectures might be applied beyond aerospace medicine, perhaps in predictive ophthalmic care on Earth, from glaucoma progression modelling to early detection of neuro-ophthalmic disease.
ReportLike0