Understanding patient experience from online medium

Improving patient experience at hospitals leads to better health outcomes. To improve this, we must first understand and interpret patients' written feedback. Patient-generated texts such as patient reviews found on RateMD, or online health forums found on WebMD are venues where patients post about their experiences. Due to the massive amounts of patient-generated texts that exist online, an automated approach to identifying the topics from patient experience taxonomy is the only realistic option to analyze these texts. However, not only is there a lack of annotated taxonomy on these media, but also word usage is colloquial, making it challenging to apply standardized NLP technique to identify the topics that are present in the patient-generated texts. Furthermore, patients may describe multiple topics in the patient-generated texts which drastically increases the complexity of the task. In this thesis, we address the challenges in comprehensively and automatically understanding the patient experience from patient-generated texts. We first built a set of rich semantic features to represent the corpus which helps capture meanings that may not typically be captured by the bag-of-words (BOW) model. Unlike the BOW model, semantic feature representation captures the context and in-depth meaning behind each word in the corpus. To the best of our knowledge, no existing work in understanding patient experience from patient-generated texts delves into which semantic features help capture the characteristics of the corpus. Furthermore, patients generally talk about multiple topics when they write in patient-generated texts, and these are frequently interdependent of each other. There are two types of topic interdependencies, those that are semantically similar, and those that are not. We built a constraint-based deep neural network classifier to capture the two types of topic interdependencies and empirically show the classification performance improvement over the baseline approaches. Past research has also indicated that patient experiences differ depending on patient segments [1-4]. The segments can be based on demographics, for instance, by race, gender, or geographical location. Similarly, the segments can be based on health status, for example, whether or not the patient is taking medication, whether or not the patient has a particular disease, or whether or not the patient is readmitted to the hospital. To better understand patient experiences, we built an automated approach to identify patient segments with a focus on whether the person has stopped taking the medication or not. The technique used to identify the patient segment is general enough that we envision the approach to be applicable to other types of patient segments. With a comprehensive understanding of patient experiences, we envision an application system where clinicians can directly read the most relevant patient-generated texts that pertain to their interest. The system can capture topics from patient experience taxonomy that is of interest to each clinician or designated expert, and we believe the system is one of many approaches that can ultimately help improve the patient experience.

Name thesis:degree_name

Ph.D.

Level thesis:degree_level

Dissertation

Discipline thesis:degree_discipline

Computer Science

Grantor

University of Illinois at Urbana-Champaign

Year dc:date

2019

Author dc:creator

• Cho, Hyun Duk

Contributors dc:contributor

• Girju, Roxana
• Zhai, Chengxiang
• Sundaram, Hari
• Paul, Michael

• healthcare
• patient-generated texts
• patient experience
• patients
• online reviews
• patient reviews
• deep neural network
• matrix factorization
• linguistics
• topic modeling
• psycholinguistics
• feature analysis

Statement dc:rights

• Copyright 2019 Hyun Duk Cho

Language dc:language

en

Handle dc:identifier

http://hdl.handle.net/2142/105760

OAI identifier oai:identifier

oai:www.ideals.illinois.edu:2142/105760