Machine learning and time-series analysis in healthcare

Lin, Yu-Wei

Permalink

https://hdl.handle.net/2142/115885 Copy

Description

Title

Machine learning and time-series analysis in healthcare

Author(s)

Lin, Yu-Wei

Issue Date

2022-07-12

Director of Research (if dissertation) or Advisor (if thesis)

Shaw, Michael J.

Doctoral Committee Chair(s)

• Shaw, Michael J.
• Seshadri, Sridhar

Committee Member(s)

• Ahsen, Mehmet Eren
• Ivanov, Anton

Department of Study

Business Administration

Discipline

Business Administration

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

Ph.D.

Degree Level

Dissertation

Keyword(s)

• Machine Learning
• Deep Learning
• Healthcare Decision Support
• Decision Support Systems
• Healthcare Data Analytics
• ICU Readmission
• Online Support Groups
• Sentiment Dynamic Features
• Knowledge Sharing
• Patient Engagement
• Behavioral Analysis
• COVID-19
• Telehealth
• Perceived Quality

Abstract

Decision support systems have altered the way consumers and businesses interact. In this dissertation, I would like to answer the question, “What is the role of decision support systems (DSS) in shaping the relationship between consumers and businesses?” In this dissertation, I focus on the impacts of DSS in healthcare. I first provide an overview of the potential and challenges of machine learning for healthcare decision support. From there, I use three Chapters to examine and empirically test the potential solutions to address the challenges of machine learning-based healthcare decision support systems. The first chapter addresses a recurring problem in the ICU administration, namely the unplanned readmission of a hospitalized patient. This chapter highlights the ability of machine learning models to improve our ICU decision-making accuracy and is a real-world example of precision medicine in hospitals. In particular, this chapter tackles the three main challenges of machine learning-based healthcare DSS, which are (1) data complexity, (2) decision criticality, and (3) model explainability. I use comprehensive, longitudinal clinical data from the MIMIC-III to predict the ICU readmission of patients within 30 days of their discharge. First, I tackle the data complexity issue by adopting dimension reduction techniques on patients’ medical records to integrate patients’ chart events, demographics, and ICD-9 code. Second, to address the decision criticality issue, I have performed in-depth deep learning performance analysis, as well as the analysis of each feature contribution to the predictive model. Third, to unpack the model explainability issue, I illustrated the importance of each input feature and their combinations in the predictive model. In this paper, I provide data-driven solutions that have strong potential to augment clinical decision-making for physicians and ICU specialists. The second chapter focuses on investigating the impact of patients’ sentiment dynamic features (SDF) on patients’ knowledge-sharing behaviors in OSGs. SDF are the patterns in which individuals’ sentiment accumulated over a period of time. I first study the impact of patients’ SDF in OSGs on their willingness to share. The results suggest that patients’ sentiment positivity, stability, and dispersion over a period of time positively affect their subsequent willingness to share in OSGs. I then analyze what affects these three SDF. I find that patients’ SDF can be influenced by their (1) online interactions and (2) offline life events. To test my hypotheses, I collect data from one of the large U.S.-based online cancer support communities about prostate cancer, which include 451 active patients with 48,692 posts and 238,846 replies. Moreover, I provide guidance for incorporating these findings into the design of OSG platforms. Further, my third chapter aims to empirically examine the relationships underlying the adoption of telehealth, patient-perceived quality, and specialty risk categories driven by COVID-19. Drawing upon computer-mediated theory and the hyperpersonal model, I explore and identify the mechanism of the positive influence of telehealth adoption. I also find that different strategies of telehealth adoption (i.e., doctors who adopt telehealth and only accept video visits, doctors who adopt telehealth and accept both video visits and in-person visits, and doctors who do not adopt telehealth) are the driving force for the improvement of patient's perception quality. Moreover, I also identify that the adoption of telehealth is more beneficial for a doctor with high-risk specialties than in low-risk specialties during the emergence phase of the pandemic outbreak. To test my hypotheses, I collect data from Zocdoc.com, an online physician appointment-booking and review website. The dataset contains 13,608 physicians with 135 specialties. One of the key managerial implications of this paper is that it provides guidance for healthcare providers to what kind of telehealth adoption model they should use during the pandemic, either the video-visit-only model or the hybrid model.

Graduation Semester

2022-08

Type of Resource

Thesis

Copyright and License Information

Copyright 2022 Yu-Wei Lin

Owning Collections