Body:

Title:  Generalizable Models for Prediction of Physiological Decompensation from Multivariate and Multiscale Physiological Time Series using Deep Learning and Transfer Learning Techniques

Committee:

Dr. Shamim Nemati, UCSD, Chair , Advisor

Dr. David Anderson, ECE, Co-Advisor

Dr. Omer Inan, ECE

Dr. Zsolt Kira, IC

Dr. Atul Malhotra, UCSD

Dr. Ravindra Metra, UCSD

Abstract:

The goal of this thesis is to develop generalizable machine learning models for early prediction of physiological decomposition from multivariate and multiscale physiological time series data. A combination of recent advances in machine learning and the increased availability of more granular physiological time series data (due to increased adoption of electronic medical records in US hospitals) has encouraged the development of more accurate prediction models for the critically ill patients. One such physiological decomposition prediction task we consider in our work is the early prediction of onset of sepsis. Sepsis is a syndromic, life-threatening condition that arises when the body's response to infection injures its own internal organs. While there are effective protocols for treating sepsis (e.g. administration of broad-spectrum antibiotics, Intravenous fluids, and vasopressors) once it has been diagnosed, there still exists challenges in reliably identifying septic patients early in their course. The purpose of this work is to explore the feasibility of utilizing low-resolution electronic medical record data and high-resolution physiological time series data to develop accurate prediction models for onset of sepsis in critically ill patients. To achieve this objective - We first investigate the connection between heart rate (HR) and blood pressure (MAP) time series - as captured through quantification of the structure of their corresponding network representation - for early signs of sepsis. We will then explore the utility of recurrent neural network models for accurate prediction of onset of sepsis. Finally, we combine ideas from adversarial domain adaptation, representation learning and conformal prediction to develop a generalizable prediction model that can adapt well to new target populations (without the requirement of obtaining gold-standard labels).