Predicting Spaceflight Associated Neuro-ocular Syndrome Using Artificial Intelligence

Abstract

Purpose : Spaceflight associated neuro-ocular syndrome (SANS) is associated with structural and functional ocular changes (e.g., disc edema) and develops in ~2/3 of astronauts during long-duration spaceflight. NASA has collected optic nerve head (ONH) optical coherence tomography (OCT) from International Space Station astronauts (in-flight data) and from the same individuals on Earth prior to space flight (pre-flight data). Our goal is to train artificial intelligence (AI) models using pre- and in-flight OCT to predict a binary outcome (SANS vs. non-SANS) using only pre-flight OCT. To address the limited amount of data, we evlauted data augmentation and a transfer learning approach using models pre-trained to identify glaucoma.

Methods : OCT imaging (Spectralis, Heidelberg Engineering) was collected from 36 astronauts and included 79 pre-flight volumes (1,392 total B-scans) and 191 in-flight volumes (3,012 total B-scans). Astronauts were labeled “SANS” if they exhibited >20 µm increase in total retinal thickness in the 250 µm segment from the Bruch’s membrane opening. Otherwise, they were labeled as “non-SANS.” There were 28 SANS astronauts and 8 non-SANS astronauts. Data was randomly separated into train (n=24), validation (n=4), and test (n=8) data by astronaut (i.e., no astronaut was in both the train and test data). ResNet50 models were trained using both pre-flight and in-flight data to predict SANS from individual B-scans and evaluated based on predicting SANS using only pre-flight data. The impact of data augmentation and pre-training were quantified. Class activation maps (CAMs) were used to identify informative OCT regions.

Results : The best performing model used both data augmentation and pre-training on a glaucoma dataset to achieve an area under receiver operating characteristic curve (AUC) of 0.86 in predicting SANS using only pre-flight data. Data augmentation had a larger impact on AUC than pre-training on the glaucoma dataset. CAMs indicated that the model focused on retinal nerve fiber layer and Bruch’s membrane to predict SANS.

Conclusions : Even with limited data, AI models show substantial promise in predicting SANS outcomes using only pre-flight OCT information. The AI models could also help to identify novel associations between ONH OCT structure and SANS.

This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.

 

Figure 1: (A) ROC curve for SANS prediction. (B) Input image along with heat map highlighting areas that influenced model predictions.