Meta-learning approach for ALS detection using sustained vowel phonations

A form of Amyotrophic Lateral Sclerosis (ALS) called Bulbar Amyotrophic Lateral Sclerosis also known as Bulbar-onset ALS, affects the bulbar region of the brain and spinal cord. The bulbar region controls essential functions such as speech, swallowing, and breathing. In bulbar ALS, the motor neurons responsible for these functions degenerate and gradually lead to muscle weakness and loss of control in the face, throat, and upper body. Currently, there is no specific cure or targeted therapy available for Bulbar ALS or any form of ALS. The effective early detection or classification of bulbar ALS using AI techniques like ML and DL necessitates a significant volume of data. This paper gives a method of speech evaluation for an automated system that distinguishes between healthy individuals and ALS subjects. The analysis of the sustained phonation vowels /a/ and /i/ in this study is specifically focused on performing automated categorization of ALS subjects. And a variety of acoustic characteristics were analysed such as Mel-frequency cepstral coefficients (MFCCs), Spectral Centroid, etc. The proposed work consists of a novel approach to the classification of Bulbar ALS using Sustained Vowel Phonations (SVP) which achieves promising performance using machine learning approaches and also using a small train dataset with a Few-Shot Learning (FSL- a paradigm of meta-learning learning) approach. The model utilizes Mel spectrograms as the input image data, which is derived from SVP. This study primarily focuses on model and algorithm, which are based algorithms among the three perspectives (data, model, and algorithm) used by few-shot learning (FSL) methods to address the few-shot problem. Experimental results demonstrate that the most successful FSL model: Prototypical networks, Matching networks, attained 86.77%, 84.8% accuracy, respectively. The efficiency of the proposed method is demonstrated with experiments on a dataset consisting of 128 SVP audio files of 64 subjects, which were collected in the Republican Research and Clinical Center of Neurology and Neurosurgery (Minsk, Belarus).