Speech Non-Linear Multiband-Time-Series Analysis for Detecting Alzheimer&rsquo;s Disease

Martim Silva; Pedro Ribeiro; Bruno Bispo; Pedro Rodrigues

Previous Article in event

A Robust Approach for Emotional Assessment: The Employment of Power-Normalized Cepstral Coefficient and Stacked Classifiers

Next Article in event

Development of intraoral sensors for the continuous measurement of clinically relevant parameters in the oral cavity

Next Article in session

Transgenic bioenergy crops as sustainable substrates for recovery of cellulosic sugars and lipids using natural deep eutectic solvents (NADESs)

Speech Non-Linear Multiband-Time-Series Analysis for Detecting Alzheimer’s Disease

^{*

1},

¹,

²,

¹ Universidade Católica Portuguesa, CBQF – Centro de Biotecnologia e Química Fina – Laboratório Associado, Escola Superior de Biotecnologia, Rua de Diogo Botelho 1327, Porto, 4169-005, Portugal.
² Department of Electrical and Electronic Engineering, Federal University of Santa Catarina, Florianópolis - SC, 88040-370, Brazil.

Academic Editor: Andrea Cataldo

Published: 11 October 2024 by MDPI in The 1st International Online Conference on Bioengineering session Biosignal Processing

Abstract:

Introduction: Alzheimer’s Disease (AD) is a prevalent neurodegenerative disorder, anticipated to triple in cases by 2050. It constitutes 50-75% of dementia cases and currently lacks a cure. Early diagnosis is crucial, allowing for treatments that may delay its progression. Traditional diagnostic methods, though effective, are invasive and expensive. Speech signal analysis has emerged as a promising non-invasive, cost-effective alternative for early AD diagnosis.

Methods: This study investigates the application of non-linear analysis under a Discrete Wavelet Transform (DWT) of speech signals for detecting AD stages. The dataset comprises 360 audio recordings from the DementiaBank Spanish Ivanova Corpus, categorized into AD, Mild Cognitive Impairment (MCI), and healthy control groups. The 360 speech signals were cleaned by removing artifacts through a filter and moments of silence utilizing Voice Activity Detection (VAD). A 50% overlap rectangular sliding window process of a 5-second duration was used, and within each window, the signal was decomposed by DWT into six bands. From each band, 10 non-linear parameters analyze the complex dynamics of our speech signals. Each feature time series is compressed over time per band, utilizing six compression metrics, and the resulting data are divided into groups based on gender and AD stage. Classical machine learning classification was implemented, and an iterative application of various normalization, feature selection, and optimization techniques was employed. The final step tested 20 classifiers to determine the most effective model for discrimination between groups.

Results and Discussion: Our findings show a 100% accuracy between men with AD and women with AD, healthy men and women with AD, and men with AD and healthy women. Furthermore, nearly all of our 15 group comparisons have an accuracy of higher than 90.9%.

Conclusion: In conclusion, our techniques culminated in a model that achieved good model performance and could differentiate between men and women, and between the three studied stages of AD.

Keywords: Alzheimer's disease, Machine learning, Non-linear Analysis, Speech signal, Wavelet transform

0 Reads
0 Recommendations