Gait force-based multimodal fusion architectures with deep neural network for gait disorders in older adults

• Main Author: Kazi Ashikur, Rahman
• Format: Thesis
• Language: English; English
• Published: 2025
• Subjects: Q Science; QA Mathematics
• Online Access: http://eprints.utem.edu.my/id/eprint/29318/

Gait analysis is an important way to help diagnose neurodegenerative diseases early, especially in older adults. Small changes in the way people walk can show the start of diseases like Parkinson’s Disease (PD), Huntington’s Disease (HD), and Amyotrophic Lateral Sclerosis (ALS). This study looks at the time and frequency patterns of gait signals using Continuous Wavelet Transform (CWT). This method helps better understand the vertical Ground Reaction Force (vGRF) signals and find small problems in walking. Even though this is useful, it is still hard to tell these diseases apart because their walking features can be similar, and aging also changes how people walk. The study used data from 64 people: 13 with ALS, 15 with PD, 20 with HD, and 16 healthy controls. Since the focus is on older adults, only data from those aged 50 and above were used. The main goal was to build a strong model that uses time-frequency features from vGRF signals along with clinical information to improve early detection and work well for different people. The work was done in three steps. First, common machine learning methods, Support Vector Machine (SVM), Random Forest (RF), Decision Tree (DT), and Multilayer Perceptron (MLP) were trained using handpicked features from the vGRF data. SVM gave the best accuracy of 83.3%. These methods worked well but depended on features chosen by hand, which limits how well work on new data. Next, deep learning methods were used by changing the vGRF signals into time-frequency images using CWT. Convolutional Neural Networks (CNN) and ResNet-50 models learned features automatically from these images. They reached accuracy rates of 95.18% and 95.06%, respectively. But these models only used signal data and did not include clinical details like age, gender, and disease severity, which can affect walking. Finally, a combined deep learning model was made that used both the spectrogram images and clinical data together. This model could learn walking patterns and personal differences at the same time. It had the highest accuracy of 98.46%, with good sensitivity, specificity, precision, and F1-score. This shows it can reliably tell apart different neurodegenerative gait disorders.