- Electrical Engineering and Computer Science
Expected date of graduation:
There are major practical and technical barriers to understanding human health, and therefore a need for methods that thrive on large, complex, noisy data. In this work, we present machine learning methods that distill large amounts of heterogeneous health data into latent state representations. These representations are then used to estimating risks of poor outcomes, and response to intervention in multivariate physiological signals. We evaluate the reduced latent representations by establishing their predictive value in important clinical tasks. Note that we value the latent space representations themselves as they provide critical insight into underlying systems. In particular, we focus on case studies that can provide evidence-based risk assessment and forecasting in settings with guidelines that have not traditionally been data-driven.
The widespread adoption of electronic health records allows us to ask evidence-based questions about the need and benefit of specific interventions in critical-care settings across large populations. Importantly, the vast amounts of data that are collected in ICUs—vital signs, clinical notes, fluids, medications—suggest an opportunity for more data- driven decision-making. Whereas clinicians may struggle to track multiple signals from multiple, rapidly evolving patients at once, algorithms excel at processing large streams of data. Computational tools that summarize relevant parts of these data streams could allow clinicians to focus on decision-making rather than just keeping up with the data.
In this thesis we propose several methods to create patient representations in a clinical setting, and use these features to predicting important outcomes. Representation learning can be thought of as a form of phenotype discovery, where we attempt to discover spaces in the new representation that are markers of important events. We argue that these latent representations are valid markers when they 1) create better prediction results on outcomes of interest, and 2) do not duplicate features that are currently known bio-markers. We explore both models that implicitly capture time with feature space concatenation, and those that explicitly model time.
In the future, we believe this work can be used to suggest therapy paths in clinical settings. While we could have simply build discriminative classifiers to target classification accuracy for any specific targeted clinical event, we believe that there is additional value in estimating underlying latent states, as this gives more information to guide choices for treatments and interventions. These estimates are also low-dimensional representations of the original signals, and may be used as a starting point for additional classification purposes on a targeted basis.