Author: Ke, Xiaoquan
Title: Automatic selection of spoken language biomarkers for dementia detection
Advisors: Mak, M. W. (EEE)
Degree: Ph.D.
Year: 2024
Subject: Dementia -- Diagnosis
Language disorders
Biochemical markers
Dementia -- Patients -- Communication -- Testing
Hong Kong Polytechnic University -- Dissertations
Department: Department of Electrical and Electronic Engineering
Pages: xx, 130 pages : color illustrations
Language: English
Abstract: Dementia is a severe cognitive impairment that may affect older adults’ health and daily lives and burden their families and caretakers. The most common form of de­mentia is Alzheimer’s disease (AD). Currently, dementia can be diagnosed through brain imaging, identification of apolipoprotein E genotypes, measuring the level of brain-derived neurotrophic factors, cerebrospinal fluid exams, and other laboratory measures. However, these measures are invasive and costly. As dementia also mani­fests itself as spoken language deficits, effective detection of early signs of the disease through the analyses of spoken languages can facilitate timely intervention to slow deterioration. This thesis analyzes a diverse set of features extracted from spoken languages and selects the most discriminative ones for dementia detection. We refer to these features as spoken language biomarkers of dementia.
This thesis proposes two deep-learning-based feature ranking (FR) methods, called dual dropout ranking (DDR) and dual-net feature ranking (DFR), to rank and select features. DDR and DFR are based on a dual-net architecture that performs feature selection (FS) and dementia detection by two neural networks: Operator and Selec­tor. The two networks are alternatively and cooperatively trained to optimize the performance of both FS and dementia detection. Specifically, in DDR, the operator is trained on features obtained from the selector to reduce classification loss, and the selector is optimized to predict the operator’s performance based on automatic regu­larization. DDR ranks features according to the probabilities that the corresponding features should be purged (or kept). In DFR, the selector is trained to find multi­ple subsets of features to predict the operator’s performance, and the operator uses these feature subsets to minimize classification errors. DFR uses all of the selector’s parameters to determine the contributions of individual features to the selector’s pre­dictions, taking into account the non-linear relationship between the input variables and the network’s output. It allows for evaluating the contributions of individual input variables in a multi-layer neural network with non-linear activation functions. We also proposed a two-step FS approach that utilizes filter methods to pre-screen features and applies more expensive FS methods to rank the pre-screened features.
The proposed FR methods were evaluated on three dementia datasets – ADReSS, AD2021, and JCCOCC-MoCA. Results on ADReSS and AD2021 show that the full feature set comprises many redundant features and that feature ranking can improve the accuracy of dementia detection. In particular, using the most dis­criminative features discovered by DDR, we achieved an F1 score of 90.4% on the ADReSS test set, which surpasses the official baseline performance by 15.9 percent­age points. Similarly, using the most discriminative features discovered by DDR, we achieve an F1 score of 86.7% on the AD2021 test set, surpassing the official base­line performance by 8.1 percentage points. The evaluations on the JCCOCC-MoCA dataset show that DFR can significantly reduce feature dimensionality while identi­fying small feature subsets with performance comparable or superior to the whole feature set. The selected features have been uploaded to https://github.com/ kexquan/AD-detection-Feature-selection, and codes are aviable at https:// github.com/kexquan/dual-dropout-ranking and https://github.com/kexquan/ dual-net-feature-ranking.
Rights: All rights reserved
Access: open access

Files in This Item:
File Description SizeFormat 
7395.pdfFor All Users5.76 MBAdobe PDFView/Open


Copyright Undertaking

As a bona fide Library user, I declare that:

  1. I will abide by the rules and legal ordinances governing copyright regarding the use of the Database.
  2. I will use the Database for the purpose of my research or private study only and not for circulation or further reproduction or any other purpose.
  3. I agree to indemnify and hold the University harmless from and against any loss, damage, cost, liability or expenses arising from copyright infringement or unauthorized usage.

By downloading any item(s) listed above, you acknowledge that you have read and understood the copyright undertaking as stated above, and agree to be bound by all of its terms.

Show full item record

Please use this identifier to cite or link to this item: https://theses.lib.polyu.edu.hk/handle/200/12961