Recurrent Multitask-Learning for Irregular Clinical Time Series Forecasting

Inflammatory Bowel Disease (IBD) is a group of chronic gastrointestinal disorders that are difficult to treat. Having no known cure, treatment courses can be long-term and expensive. IBD flare-ups can happen without warning and there exists no objective criteria to measure the disease's activity. Recently, Recurrent Neural Networks (RNN) have emerged as a state-of-the-art method in clinical time series analysis; building on recent work that apply RNNs to temporal patient data, this thesis explores methodologies for processing temporal clinical data, the feasibility of a deep RNN classifier to forecast the future healthcare utilization, and techniques to curb overfitting while training on a small dataset. This work shows that multitask learning is helpful to train stable models, and deep networks can be engineered to process small noisy datasets in the clinical domain.