To create deep learning artificial intelligence (AI) models for predicting the development of Spaceflight Associated Neuro-Ocular Syndrome (SANS) using optical coherence tomography (OCT) imaging of the optic nerve head.

Retrospective analysis.

AI deep learning models were trained to predict SANS onset by using two OCT datasets: pre- and inflight OCT images acquired from astronauts (flight data) and pre- and in-bedrest images from research participants undergoing head-down tilt bedrest as an Earth-bound model of SANS (ground data). Both datasets were partitioned by participant into training and testing data. Resnet50-based models were trained using exclusively flight data, exclusively ground data, and a combination of both. All models were evaluated based on their ability to predict SANS using only pre-flight or pre-bedrest imaging in both datasets. Performance was assessed using receiver operating characteristic areas under the curve (AUC). Class activation maps (CAMs) were generated to identify impactful image regions.

The model trained on flight data achieved an AUC (95% CI) of 0.82 (0.54-1.0) on flight data and 0.67 (0.51-0.83) on ground data. The ground-trained model achieved an AUC of 0.71 (0.50-0.91) on ground data and 0.76 (0.51-0.91) on flight data. The combined model achieved an AUC of 0.81 (0.53-0.95) and 0.72 (0.52-0.92) on flight and ground data, respectively. CAMs identified peripapillary regions of the nerve fiber layer, retinal pigmented epithelium, and anterior lamina surface as most important for predictions.

AI models can predict SANS based on pre-flight OCT imaging with moderate-to-high performance even in this data-limited setting. The performance of cross-trained models and similarity in CAMs suggests similarity between SANS-related changes in flight and ground datasets, proving further support that head-down tilt bedrest is a reasonable Earth-bound model for SANS. NOTE: Publication of this article is sponsored by the American Ophthalmological Society.