Ali, Nur Alisa (2024) Detection of autism spectrum disorder based on time domain electroencephalogram signal using enhanced BILSTM models. Doctoral thesis, Universiti Teknikal Malaysia Melaka.
![]() |
Text (24 Pages)
Detection of autism spectrum disorder based on time domain electroencephalogram signal using enhanced BILSTM models.pdf - Submitted Version Download (464kB) |
![]() |
Text (Full Text)
Detection of autism spectrum disorder based on time domain electroencephalogram signal using enhanced BILSTM models.pdf - Submitted Version Restricted to Registered users only Download (4MB) |
Abstract
The global prevalence of Autism Spectrum Disorder (ASD) has driven researchers to develop well-defined automated approaches for early detection, surpassing standard behavioral assessments. Behavioral assessment methods face challenges to detect ASD at early stage because pronounced symptoms of autism are often observed between the ages of two and three years old, leading to delayed or missed diagnoses in some individuals. The electroencephalogram (EEG) has emerged as a promising quantifiable tool for identifying ASD biomarkers earlier than standard behavioral assessments. Its integration with deep learning methodologies has advanced ASD diagnosis through computer-aided diagnosis (CAD) systems. This research intends to classify the time-series EEG data of ASD and typical development (TD) samples from the SFARI dataset, which comprises 53 subjects (14 TD and 39 ASD) ranging in age from 10 months to 21 years. Two deep learning methods are particularly suitable for handling time-series data, namely Convolutional Neural Networks (CNN) and Recurrent Neural Networks (RNN) family. While most researchers utilize CNN-based approaches that require conversion from the time domain to the time-frequency domain, this study explores the potential of a Long Short-Term Memory (LSTM)-based models from RNN family to classify EEG data directly in the time domain. Specifically, this research examines the efficacy of LSTM and Bidirectional Long Short-Term Memory (BiLSTM) networks in distinguishing between ASD and TD individuals without relying on prior demographic knowledge or the requirement of data conversion. The optimized BiLSTM model achieved an accuracy of 99.68%, outperforming the LSTM in classifying ASD and TD using 117 multichannel EEG recordings. However, managing multichannel EEG data presents challenges, particularly with unpredictable ASD individuals. To address this, a hybrid model incorporating Autoregressive (AR) feature extraction, General Learning Equilibrium Optimizer (GLEO) feature selection, and optimized BiLSTM was developed to perform channel selection. This hybrid method achieved 99.89% accuracy using only 29 EEG channels, thereby reducing the complexity of the experimental setup by 75%. The discriminative ability of each channel in distinguishing between ASD and TD EEG data was supported through a one-way Analysis of Variance (ANOVA) method. This analysis revealed that 27 channels produced significantly different outputs, while the remaining 2 channels yielded p-values slightly higher than 0.05. These findings underscore the reliability of the proposed AR-GLEO-BiLSTM method for diagnosing ASD and lay a foundation for detecting ASD biomarkers in individuals before behavioral diagnosis is typically possible, or when behavioral features are not apparent until two years of age or later.
Item Type: | Thesis (Doctoral) |
---|---|
Uncontrolled Keywords: | Autism Spectrum Disorder (ASD), Early detection, Behavioral assessment |
Divisions: | Library > Tesis > FTKEK |
Depositing User: | Muhamad Hafeez Zainudin |
Date Deposited: | 17 Mar 2025 12:01 |
Last Modified: | 17 Mar 2025 12:01 |
URI: | http://eprints.utem.edu.my/id/eprint/28564 |
Statistic Details: | View Download Statistic |
Actions (login required)
![]() |
View Item |