Identifying Disease Subtypes from Complex Medical Data

Many contemporary fields of medicine produce increasingly high-dimensional data. "Traditional" data analysis methods trained in clinical research education are insufficient to investigate such highly complex data. Therefore, various machine learning techniques are increasingly finding their way into medical data analysis. However, disease diagnosis itself has traditionally been defined on simple rules based on easily observable low-dimensional data. Traditional clinical diagnoses can therefore be inappropriate as ground truths to train machine learning classifiers. Here, using the eye disease of glaucoma as an example, we demonstrate how patterns identified by unsupervised machine learning on complex medical data can help to refine and quantitatively re-define disease diagnosis. Our patterns of vision loss due to glaucoma enable a whole field of quantitative research which was not available by previously existing diagnostic approaches.