Date of Award
Spring 2023
Document Type
Dissertation
Degree Name
Doctor of Philosophy (PhD)
Department
Electrical Engineering (ENAS)
First Advisor
Tagare, Hemant
Abstract
Measuring symptoms over time is an essential part of clinical research. Differencesin symptom progression are often used to better understand diseases and identify treatment targets. In clinical trials, change in symptoms is always a key finding. Many symptoms, however, are often difficult to measure empirically, particularly in neurological and psychiatric conditions. In these cases, clinical questionnaires which measure symptoms on ordinal scales are used. When longitudinal clinical studies use these questionnaires, noisy ordinal time series are created. Current approaches to the analysis of these questionnaires do not properly account for the noise or the ordinality and often all questions are summed together to create a total score. As a result, information about the progression of specific symptoms is lost that may shed light on the heterogeneity of a disease. In this dissertation, we bring sophisticated machine learning techniques to the analysis of these time series. First, we present a novel model for denoising ordinal time series. We discuss the challenges of denoising clinical ordinal times series and show how our model is able to address them. Our model is based on standard ordinal regression techniques to preserve the ordinal properties and utilizes a hierarchical prior to share information across the population. We validate the performance of our hierarchical model on simulations of noisy and heterogeneous time series. Although our ordinal model can be applied generally, we focus on denoising ordinal scores of Parkinson’s disease motor symptoms. Parkinson’s disease heterogeneity is not yet well understood and clinical ordinal questionnaires are a key metric in this research. We show that our denoising model can elucidate motor symptom progression in individuals with Parkinson’s disease. We present multivariate analysis techniques to examining the motor symptoms scores on a per-question basis. This analysis is only made possible by denoising. There are many common assumptions in previous research on Parkinson’s disease heterogeneity. We show how techniques such as dimensionality reduction and hierar- chical clustering can be used to test these assumptions. We assess the relationship between motor symptom progression and a range of other disease factors, including genetics and brain imaging. These results are used to explain open questions and inconsistencies in the existing literature about Parkinson’s disease. Lastly, we show that our methodologies can be easily adapted for use with other diseases and study designs. To do this, we apply our novel denoising model to a clinical study of depression comparing a treatment to a placebo. We verify the advantages of denoising in this new application and compare the placebo and treatment groups after denoising. We also explore how this model can be used for the prediction of ordinal time series.
Recommended Citation
Koss, Jonathan D., "Machine Learning for Ordinal Time Series to Understand Disease Progression" (2023). Yale Graduate School of Arts and Sciences Dissertations. 1094.
https://elischolar.library.yale.edu/gsas_dissertations/1094
