Deep Learning Models for Clinical Data: Addressing Task Specific Structure

As the adoption of electronic health records (EHRs) increases, so do the opportunities to improve patient care using these data. Applied to high-dimensional EHR data, machine learning techniques can help identify complex relationships between patient covariates and outcomes. However, in order to augment clinical care these models must generalize to (i.e., perform well on) never-before-seen data. In healthcare settings, generalization performance is often hindered by limited training data. Though clinical data are high dimensional, there is often a limited number of examples that can be used for training, due to low incidence rates of the outcomes of interest. To address this challenge, we develop and evaluate methods that combine deep learning techniques with knowledge about the task structure to improve sample efficiency. Throughout this dissertation, we augment learning algorithms by exploiting known task structure pertaining to i) invariances, ii) signal dynamics, and address challenges associated with iii) class imbalance driven low homophily. First, different types of temporal invariances (e.g., phase invariance) are present in clinical time-series tasks. However, such invariances may vary across tasks. We propose a novel approach, `Sequence Transformer Networks' that learns to recognize and exploit task-specific invariances, reducing intra-class variance and improving performance. Second, risk factors for a given adverse outcome may change as a patient's admission progresses. Though techniques like recurrent neural networks (RNNs) should in theory be able to capture such time varying dynamics, when training data are limited performance can suffer. We propose a novel RNN-based architecture in which we textit{relax weight sharing} over time to capture time-varying relationships. Finally, graph neural networks (GNNs) are a popular method for learning feature representations from graphs but are often evaluated on tasks with graphs presenting high homophily (or high similarity among connected nodes). In fact, GNNs are known to do poorly in low homophily settings. In clinical tasks, high class imbalance leads to asymmetrical homophily: high homophily with respect to the majority class and on average, but low homophily with respect to the minority class of interest. We address class imbalance driven low homophily by evaluating an attention-based mechanism against recently proposed methods for dealing with low homophily in general. By adapting techniques to better leverage task structure such as class imbalance driven low homophily, we are able to improve sample efficiency and predictive performance on tasks such as estimating the risk for adverse patient outcomes.PHDComputer Science & EngineeringUniversity of Michigan, Horace H. Rackham School of Graduate Studieshttp://deepblue.lib.umich.edu/bitstream/2027.42/171373/1/jeeheh_1.pd