Thesis
AGTD - The AudioGene Translational Dashboard: a hybrid machine learning and visualization interface for genetic diagnosis of Autosomal Dominant Non-Syndromic Hearing Loss
University of Iowa
Master of Science (MS), University of Iowa
Spring 2024
DOI: 10.25820/etd.007493
Abstract
Autosomal Dominant Non-Syndromic Hearing Loss (ADNSHL) presents a major genetic diagnosis challenge due to the complex heterogeneity of the molecular basis of the disease – i.e., multiple distinct gene mutations. This thesis introduces a new approach to diagnosis that combines ensemble and semi-supervised Support Vector Machine (SVM) models with advanced visualization tools to improve the accuracy and confidence of ADNSHL diagnoses. Previous methods explored were AudioGene V4 (AG4), an MI-SVM Classifier, and AudioGene 9/9.1 (AG9 and AG9.1), an Ensemble Classifier, developed by Kyle Taylor (AG4), Chibuzo Nwakama (AG9), and Sean Ryan (AG9.1) respectively. These machine-learning algorithms were insufficient to combat the class imbalance and the lack of representation of the population within our training dataset. This led to poor accuracy within the MI-SVM and ensemble models, with 44% and 47% respectively in LOOCV. Therefore, using our new novel approach of integrating the ensemble and MI-SVM models with an advanced visualization dashboard as part of the AudioGene website. The aim is to enable clinicians to dynamically interact with the model predictions, allowing the clinician to see inside the “Black Box” of AudioGene models and help guide the clinicians in customizing the model to make better predictions. The visualization dashboard features an interactive three-dimensional audiogram plot, a 2-d audio profile plot, a surface profile viewer, and an ethnicity distribution visualization. Our evaluation through case studies reveals that AGDT can enhance diagnosis confidence and improve interpretation of AudioGene predictions. This thesis concludes that visualization tools are helpful in guiding the clinician in understanding AudioGene models and providing a promising advancement in ADNSHL diagnosis. More generally, it demonstrates the power of hybrid AI-human expert systems in delivering explainable AI.
Details
- Title: Subtitle
- AGTD - The AudioGene Translational Dashboard: a hybrid machine learning and visualization interface for genetic diagnosis of Autosomal Dominant Non-Syndromic Hearing Loss
- Creators
- Benjamin DeSollar
- Contributors
- Thomas Casavant (Advisor)Kishlay Jha (Committee Member)Terry Braun (Committee Member)
- Resource Type
- Thesis
- Degree Awarded
- Master of Science (MS), University of Iowa
- Degree in
- Electrical and Computer Engineering
- Date degree season
- Spring 2024
- Publisher
- University of Iowa
- DOI
- 10.25820/etd.007493
- Number of pages
- xi, 89 pages
- Copyright
- Copyright 2024 Benjamin DeSollar
- Language
- English
- Date submitted
- 04/19/2024
- Description illustrations
- Illustrations, tables, graphs, charts
- Description bibliographic
- Includes bibliographical references (pages 87-89).
- Public Abstract (ETD)
This research explores a new strategy to tackle the diagnostic complexities and challenges of human deafness (ADNSHL). Previous methods explored using machine learning algorithms needed to be more robust to combat the class imbalance and the lack of representation of the population within our training dataset. This led to poor accuracy within the multi-audiogram-trained model (AG4) and the single audiogram-trained model (AG9/AG9.1), with 44% and 47%, respectively. Therefore, we developed a novel approach to integrate AG9.1 and AG4 with an advanced visualization dashboard. We developed this hybrid method to have the clinicians guide the model to make more accurate predictions and provide statistical significance within their diagnoses. The visualization dashboard features interactive three-dimensional audiogram plots, audio profile charts, surface profile viewers, and ethnicity distribution visualizations. Our evaluation through case studies reveals that AGDT can help in the diagnostic confidence and interpretability of AudioGene’s genetic predictions. This thesis concludes that visualization tools are useful in assisting clinicians to guide the AudioGene models and provide promising advancements in ADNSHL diagnoses.
- Academic Unit
- Electrical and Computer Engineering
- Record Identifier
- 9984647256502771
Metrics
10 Record Views