Deep Semi-Supervised and Multi-Stage Learning for Medical Applications
Abstract
Machine learning techniques are widely used to build models for applications in healthcare. These models typically predict likelihood of a particular patient outcome in a given setting. For clinical utility, these models are often used to derive parsimonious models that predict outcome risks of certain populations. Training these models on a specific patient population, their demonstrated utility is confined to patients with characteristics similar to the original derivation cohort. However, these simpler machine learning techniques may lack the discriminatory power to recognize subpopulations within a population that behave or respond differently to identical interventions. Conversely, while more complex machine learning techniques and complex data streams may possess the sophistication necessary to recognize and appropriately predict outcomes of these subpopulations, the training sizes necessary to achieve good results are prohibitively large. Correctly understanding and identifying the differences and similarities that separate and unify various subpopulations is key to building a model that is sufficiently extensible to explain population variance while minimizing unnecessary complexity.
This dissertation applies and advances machine learning for healthcare through three approaches. First, it utilizes advanced machine learning techniques for clinical modeling. This is done while predicting harmful outcomes such as mortality in vulnerable patient populations. Second, it describes advanced machine learning techniques to handle heterogeneity in retrospective analyses. It develops a novel application of a deep mixture of experts to describe this heterogeneity, learning phenotypes in a risk-driven method. Finally, it describes needs and opportunities in harnessing remote sensors for health monitoring and details two specific approaches to extracting useful health data from longitudinal sensors
Citation
Hurley, Nathan Clinton (2022). Deep Semi-Supervised and Multi-Stage Learning for Medical Applications. Doctoral dissertation, Texas A&M University. Available electronically from https : / /hdl .handle .net /1969 .1 /197083.