We propose an unsupervised learning approach that can be tied back to existing metadata, like mortality, age, BMI, etc. To accomplish this, we will train an Autoencoder model to create a low-dimensional representation of each image (Bank et al. 2020), and then use different clustering methods to determine optimal groupings for these images based on their encoding (Song et al. 2013)(Guo et al. 2017). Once these groups are instantiated, we can then associate image metadata to each cluster to determine whether there are statistically significant attributes tied to specific clusters. If it could be proven that attributes like mortality rate or success with intubation are linked to certain clusters, that information could be incredibly valuable for clinical outcomes. Also, although we have limited prognosis labels, we will also determine autoencoder performance by trying to classify the image based on the encoding using fully connected layers.
