- 128 views
- 119 downloads
Alzheimer’s Dementia Detection Through Machine Learning: Analyzing Linguistic and Acoustic Features in Spontaneous Speech
-
- Author / Creator
- Shah, Zehra
-
With the rapid aging of the world’s population, the global burden of aging related mental disorders, such as Alzheimer’s dementia (AD), is also on the rise. Unfortunately global healthcare systems are vastly under-resourced, which means that many people who need mental health services are unable to receive it. There is a critical need for timely, inexpensive, objective and scalable mental health screening and monitoring methodologies to augment currently available diagnostic tools. Speech analysis has the potential to be utilized as a window into the state of the human mind, meaning it could provide support for timely, reliable, and objective screening of many psychiatric disorders including AD.
In this dissertation, we present our research on machine learned models that can diagnose AD based on linguistic and acoustic features derived from speech. We show that AD can be detected reliably from spontaneous speech samples, and that this can be done even independently of the language spoken. The first part of this thesis presents machine learned models based on linguistic and acoustic features derived from spontaneous English speech samples. We find that linguistic features alone perform well, reaching 85% balanced accuracy on a hold-out test set, and ensemble models based on linguistic and acoustic features show comparable or slightly lower accuracy. The second part presents models that use features derived from measures of speech rate, complexity and intelligibility, this time in a cross-lingual setting (training on English speech samples and testing on Greek speech samples). These learned models, despite the significant domain shift between the training and test sets, reached a relatively high balanced accuracy of 70%, showing that AD detection from speech is possible even across two different languages. Furthermore, we provide some exploratory data analyses of the features derived in the cross-lingual experimental setting, and show that some of these features have a visibly discriminating pattern that can be successfully utilized for clustering the samples.
This work paves the way for building automated machine learned systems for detecting and monitoring AD. With further validation on larger and more diverse data sets, such systems have the potential of being deployed at scale to flag early signs of AD and monitor the progression of AD severity.
-
- Graduation date
- Fall 2023
-
- Type of Item
- Thesis
-
- Degree
- Master of Science
-
- License
- This thesis is made available by the University of Alberta Libraries with permission of the copyright owner solely for non-commercial purposes. This thesis, or any portion thereof, may not otherwise be copied or reproduced without the written consent of the copyright owner, except to the extent permitted by Canadian copyright law.