Finding disease subtypes by ignoring what healthy people share

When doctors search for disease subtypes, noise from normal human variation gets in the way. This work uses contrastive learning to identify patient subgroups driven purely by disease factors, ignoring common patterns shared with healthy controls. The method uses a deep learning model that optimizes a custom loss function via expectation-maximization, tested on MNIST and four medical imaging datasets with improved results over prior approaches. Code and datasets are released.