Dynamic Machine Learning using Signal Processing and Tensor-Based Methods to Predict Clinical Outcomes

Abstract

Machine learning and artificial intelligence can be used in improving patient care by providing important insights from the data generated throughout the duration of a patient's stay. Routinely collected data, such as electrocardiogram and electronic health record data, are two such examples of data that are frequently recorded in hospital settings. Electrocardiogram, specifically, is a noninvasive and continuously updated measure of a patient's cardiac electric activity, and as such, has the potential to provide a real-time view of a patient's current status.

This research is composed of four projects. In the first, we propose a system of signal processing for both heart rate variability and electrodermal activity to detect poor sleep quality of people with fibromyalgia, using a wearable device. In the second, we introduce a framework of processing signals using both Taut String and tensor decomposition to (1) extract meaningful features from input signals and (2) reduce the feature space to only the most pertinent information, while maintaining structural information from the input signals. This framework is applied to three cohorts of patients from Michigan Medicine, with each cohort increasing in heterogeneity. The physiological signals and electronic health record information collected from the patients in each cohort were used to predict adverse outcomes post-surgery. This study serves as validation to previous work on post-cardiac surgery, as well as generalizing the methodology outwards to other types of surgery. The third project uses the framework developed in the second for patients in the intensive care unit at risk to develop sepsis. The goal is, using continuous electrocardiogram, arterial line, and/or electronic health record data, to predict which patients are at more risk to develop poor outcomes related to sepsis. The fourth project expands upon the third by incorporating the learning using privileged information paradigm into the same sepsis prognosis design.

Together, the four projects discussed in this thesis contribute to dynamic trajectory prediction using signal processing in different health contexts. These studies demonstrate that further study of electrocardiogram data's utility in clinical decision support systems is warranted.