ORCID
Kylar Akira DeLoach: https://orcid.org/0009-0000-5417-2025
Degree
Bachelor of Science (B.S.)
Major(s)
Computer Science
Document Type
Immediate Open Access
Abstract
Alzheimer's disease (AD) is a growing global health concern, with millions of people affected worldwide and cases expected to rise significantly in the coming decades. Early detection is critical for patient treatment and care, and recent advances in natural language processing (NLP) have shown promise in identifying linguistic markers associated with AD. However, most existing work has focused on English, leaving speakers of other languages with limited access to such tools. This study investigates how effective AD detection models trained on English data are at transferring to Greek, a low-resource language with limited dementia-related speech data available. We propose a cross-lingual transfer framework that evaluates two complementary paradigms: input-space alignment using monolingual large language models (LLMs), and target-space generalization using multilingual LLMs. Both paradigms are evaluated across three transfer strategies: standard monolingual training, zero-shot transfer, and machine translation, with an additional bidirectional translation experiment applied to the monolingual LLMs. All models were trained on English transcripts from the Cookie Theft picture description task and evaluated on both English and Greek transcripts from the same task. Our results show that the effectiveness of cross-lingual transfer depends on both model type and size. Multilingual LLMs demonstrated greater robustness on native low-resource language data, while smaller monolingual LLMs maintained or improved performance through translation-based pipelines. Larger monolingual models and bidirectional translation consistently showed performance degradation, suggesting that single translation pipelines paired with smaller monolingual models offer a viable pathway for low-resource AD detection. These findings suggest that NLP-based AD detection can be extended to low-resource languages, though further work is needed before such systems could be used reliably in clinical practice.
Date Defended
4-30-2026
Thesis Director
Dr. Nisha Pillai
Second Committee Member
Dr. Charan Gudla
Third Committee Member
Dr. Matthew Peaple
Recommended Citation
DeLoach, Kylar A., "Dementia Detection in Low-Resource Languages: Evaluating Translation-Assisted Transfer Learning for Multilingual Clinical Assessment" (2026). Honors Theses. 205.
https://scholarsjunction.msstate.edu/honorstheses/205
Rights Statement
“Dementia Detection in Low-Resource Languages: Evaluating Translation-Assisted Transfer Learning for Multilingual Clinical Assessment,” Copyright 2026 by Kylar Akira DeLoach. All rights reserved. This thesis contains original scholarly work by the author, including analyses, methodological contributions, and derived computational models. Portions of this research utilize data from the DementiaBank corpus distributed through the TalkBank system, which remain subject to their respective copyright, data-use agreements, and access restrictions. Any reuse of DementiaBank materials beyond fair use must comply with the licensing and permission requirements established by the original rights holders and the TalkBank/DementiaBank project. Readers seeking access to the underlying corpus should obtain authorization directly from the data providers.
Included in
Artificial Intelligence and Robotics Commons, Diagnosis Commons, Language Interpretation and Translation Commons, Linguistics Commons
