Vocal Clues to Diabetes Mellitus: Exploring the Ethics and Tech of AI in Clinical Practice Restricted; Files Only
Weinstein, Samuel E (Spring 2023)
Abstract
Diabetes mellitus (DM) is one of the most common diseases globally. It incurs enormous economic burdens at the personal, national, and global levels. Unfortunately, inequalities in healthcare systems between health plans and geographical regions and the high direct and indirect healthcare expenses directly impact healthcare access, especially for those in lower economic classes. These factors make access to DM screening challenging and inequitable, contributing to almost one-quarter of all people with diabetes mellitus remaining undiagnosed. Although widely accepted, current standard tests for DM are available only in specially equipped medical centers, are often painful, time-consuming, and can be cumbersome to schedule. Taken together, there is a need for non-invasive methods for diabetes mellitus screening that can be easily accessed. As a sub-study of The Voice Study, which aims to develop a non-invasive and highly accessible screening tool for diagnosing DM using the human voice, my thesis investigates whether the current data of The Voice Study would generate a biased AI algorithm by determining if a variation in voice acoustics exists between sexes and among racial subgroups. The consequences of a biased AI system, specifically in healthcare, would be detrimental as it would result in incorrect diagnoses for specific patient populations, perpetuating and potentially exacerbating already existing disparities in healthcare. Participants included were screened and provided informed consent. Voices were recorded in Emory Healthcare clinics and analyzed using a Computerized Speech Lab with Multi-Dimensional Voice Program software. The 34 parameters of each voice record were analyzed within the subgroups the participant identified. The results display significant variations in multiple parameters between male and female voices. Acoustic variation was also found among racial groups in both male and female subgroups. These results indicate that attention should be given to the diversity of data when designing and deploying AI algorithms, especially in clinical practice. This study should be advanced with more data and analyzed together with other studies to develop specific recommendations for tackling biases in diagnostic AI algorithms, as well as others deployed in clinical practice.
Table of Contents
Preface 1
Introduction 2
Literature Review 6
Physiological influence of diabetes mellitus on voice 6
Acoustic variation among demographic populations 11
Ethical considerations of artificial intelligence in healthcare 15
Methods 18
A. Voice Study Recruit 18
Participants Protocol
Glucose measurements
Voice recording
Voice acoustic assessment
Statistical analysis
B. Demographic Regrouping 22
Results 24
Discussion 30
References 32
About this Honors Thesis
| School | |
|---|---|
| Department | |
| Degree | |
| Submission | |
| Language |
|
| Research Field | |
| 关键词 | |
| Committee Chair / Thesis Advisor | |
| Committee Members |
Primary PDF
| Thumbnail | Title | Date Uploaded | Actions |
|---|---|---|---|
|
|
File download under embargo until 19 May 2025 | 2023-04-16 16:38:53 -0400 | File download under embargo until 19 May 2025 |
Supplemental Files
| Thumbnail | Title | Date Uploaded | Actions |
|---|