Voice-Based Amyotrophic Lateral Sclerosis(ALS) Classification Using Deep Learning Method

Abstract

Amyotrophic Lateral Sclerosis (ALS) is a progressive neurodegenerative disease that severely impacts motor functions. Early identification of ALS remains a significant challenge due to the variability in its early symptoms, particularly speech and voice issues. This thesis investigates the feasibility of leveraging deep neural networks (DNNs) for automatic voice-based classification of ALS in its early stages. The study employed voice data from 20 healthy individuals, 27 ALS patients with dysarthria, and 16 ALS patients without dysarthria.Temporal features (jitter, shimmer, HNR) and time-frequency features (MFCCs) were extracted from sustained phonation of syllables and words. These features were then fed into different DNN architectures for classification. Convolutional Neural Networks (CNNs) utilizing mel-frequency cepstral coefficients (MFCCs) achieved an accuracy of 99.25% for two-class classification of ALS and healthy individuals, and 98.35% for three-class classification (ALS with dysarthria, ALS without dysarthria, and healthy individuals). This outperformed Long Short-Term Memory networks and other classical machine learning methods. This study demonstrates the potential of DNNs, particularly CNNs with MFCC features, for early and accurate ALS detection. This holds promise for improved clinical management and development of novel therapeutic interventions.