Approved research

Learning Multimodal Representations for the Prediction of Human Disease

Lay summary

Meaningful data representations have been critical to the success of recent machine learning methods. For example, the representations of images learned by convolutional networks have revolutionized the field of computer vision. Clinical data lacks such natural representations, which has limited our ability to develop machine learning models that can be used effectively and fairly in clinical practice. We focus on constructing meaningful representations from heterogeneous clinical data, using structured models to transform a patient's medical history into a continuous representation. In addition to demonstrating the value of these representations in the prediction of disease outcomes, we will identify the robustness of representations used in machine learning methods that integrate electronic health records, images, genomics and other clinical data. These advancements will additionally be useful to increasing the generalizability of machine learning models to new patients, leading to increased adoption in various healthcare settings. We request the full cohort to maximize statistical power and phenotype coverage.