Boston University
Deep learning-based modeling of digital voice for cognitive assessment
Abstract
dc:description.abstractDementia, particularly Alzheimer's disease (AD), poses a significant public health challenge, with rising prevalence and substantial societal costs. Early detection is crucial for timely intervention, yet current diagnostic methods remain invasive, costly, and inaccessible. This thesis explores the application of deep learning techniques to detect dementia using voice recordings, leveraging both acoustic and linguistic features for assessment. Voice data from the Framingham Heart Study (FHS), comprising 1,994 neuropsychological interview recordings from 917 participants across cognitive statuses—Normal Cognition (NC), Mild Cognitive Impairment (MCI), and Dementia (DE)—were used for model development. The study investigates state-of-the-art deep learning models, including Temporal Convolutional Networks (TCNs) and wav2vec2 embeddings for audio analysis and BERT-based transformer models for text analysis. Evaluation metrics, including AU-ROC, AU-PRC, F1-score etc., were used to assess model performance. Results demonstrated that text-based models outperformed audio-based models in distinguishing between NC and DE (mean AU-ROC: 0.938 vs. 0.852) and NC and MCI (AU-ROC: 0.730 vs. 0.628). These findings suggest that linguistic impairments are more pronounced in later-stage dementia, while audio features offer limited discriminatory power for early-stage detection. By modeling digital voice through advanced AI techniques, this research contributes to the development of scalable, non-invasive screening tools for early dementia detection.
Degree
thesis:*- Name thesis:degree_name
- Master of Science
- Level thesis:degree_level
- masters
- Discipline thesis:degree_discipline
- Computer Science
- Grantor
- Boston University
- Year dc:date.issued
- 2025
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Varshney, Abhishek
- Advisor dc:contributor.advisor
-
- Kolachalama, Vijaya B.
- ORCID dc:identifier.orcid
Subjects
dc:subject × 5Rights
dc:rights- Statement dc:rights
-
- Attribution 4.0 International
- Licence dc:rights.uri
- Language dc:language.iso
- en_US
Identifiers
dc:identifier.*- Handle dc:identifier.uri
- https://hdl.handle.net/2144/53425