Unveiling Key Features: A Comparative Study of Machine Learning Models for Alzheimer's Detection - Thesis

Aug 3, 2026

Upload by Eliana

This thesis rigorously evaluates the application of NLP techniques and machine learning models to identify linguistic signatures indicative of dementia using the DementiaBank Pitt corpus. A binary classification framework combines embedding methods (Doc2Vec, Word2Vec, GloVe, BERT) with traditional algorithms (Random Forest, Multinomial Naïve Bayes, ADA boost, KNN, Logistic Regression) and deep learning models (LSTM, Bi-LSTM, CNN-LSTM). Model effectiveness is assessed by performance separating dementia-affected transcribed speech from control subjects. Feature interpretability is enhanced via LIME, SHAP, permutation importance, and integrated gradients, clarifying influential variables in predictions.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages63

Reading Time2 h 39 min

AccessFree to download

Introduction

Background on Alzheimer's disease

Economic and caregiving impact

Lack of cure and symptom progression

Symptom progression across mild, moderate, severe stages

Utah State University DigitalCommons@USU | All Graduate Reports and Creative Projects, Fall
2023 to Present | Graduate Studies | | --- | --- | | 8-2024
Unveiling Key Features: A Comparative Study of Machine Learning Models for Alzheimer's Detection
Cailean Bushnell Utah State University
Follow this and additional works at: [https://digitalcommons.usu.edu/gradreports2023](https://digitalcommons.usu.edu/gradreports2023)
Part of the Business Analytics Commons | | Recommended Citation Bushnell, Cailean, "Unveiling Key Features: A Comparative Study of Machine Learning Models for Alzheimer's Detection" (2024) . All Graduate Reports and Creative Projects, Fall 2023 to Present. 42. [https://digitalcommons.usu.edu/gradreports2023/42](https://digitalcommons.usu.edu/gradreports2023/42) This Creative Project is brought to you for free and open access by the Graduate Studies at DigitalCommons@USU. It has been accepted for inclusion in All Graduate Reports and Creative Projects, Fall 2023 to Present by an authorized administrator of DigitalCommons@USU. For more information, please [contact digitalcommons@usu.edu](contact digitalcommons@usu.edu). Unveiling Key Features: A Comparative Study of Machine Learning Models for Alzheimer's Detection By: Cailean Bushnell A thesis submitted in partial fulfillment of the requirements for the degree of MASTER OF SCIENCE In Economics Unveiling Key Features: A Comparative Study of Machine Learning Models for Alzheimer's Detection Cailean Bushnell Department of Economics and Finance Utah State University Abstract This thesis rigorously evaluates the application of an array of natural language processing (NLP) techniques and machine learning models to identify linguistic signatures indicative of dementia, as sourced from the DementiaBank Pitt corpus. Utilizing a binary classification paradigm, this study meticulously integrates sophisticated embedding methods—including Doc2Vec, Word2Vec, GloVe, and BERT—with traditional machine learning algorithms such as Random Forest, Multinomial Naïve Bayes, ADA boost, KNN classifier, and Logistic Regression, alongside deep learning architectures like LSTM, Bi-LSTM, and CNN-LSTM. The efficacy of these methodologies is evaluated based on their capacity to differentiate between transcribed speech impacted by dementia and that from control subjects. To enhance interpretability, this research also employs feature importance analysis through LIME, SHAP, permutation importance, and integrated gradients, shedding light on the variables most instrumental in driving model predictions. The results of this comprehensive analysis not only illuminate the robust potential of these combined NLP and machine learning approaches in the context of medical screening but also contribute additional valuable insights to the field of NLP and dementia screening specifically. Keywords: dementia detection, machine learning, deep learning, Pitt corpus, feature importance. I acknowledge and thank Carly Fox for her guidance and for serving as my thesis committee chairman. I also thank Todd Grifith and Pedram Jahangiry for serving as thesis committee members. SECTION I: Introduction Alzheimer's disease (AD), a progressive neurodegenerative disorder that causes atrophy of the brain (the loss of neurons and connections between neurons), is the most common form of dementia, making up 60%-70% of all dementia cases [10] . AD is characterized by a decline in cognitive function and impacts the daily life of a diagnosed individual significantly. It is currently the 7th leading cause of death and a leading cause of disability within the elderly population [11] . It relentlessly robs individuals of their memories, reasoning skills, and ultimately, their independence. Currently, over 55 million individuals are living with AD, with a disproportionate percentage of that 55 million (60%), living in low-and middle-income countries [11] . The econom

FAQs

What dataset is used to evaluate the models for dementia detection?

The study sources data from the DementiaBank Pitt corpus.

Which modeling approaches are compared in the thesis?

It compares embedding techniques (Doc2Vec, Word2Vec, GloVe, BERT) with traditional machine learning classifiers and deep learning architectures such as LSTM, Bi-LSTM, and CNN-LSTM.

How does the thesis improve interpretability of model predictions?

It applies feature importance and explanation methods including LIME, SHAP, permutation importance, and integrated gradients to identify influential variables.

Aug 3, 2026

Upload by Eliana