Graph Neural Networks to Indentify Amyotrophic Lateral Sclerosis

Abstract

Amyotrophic Lateral Sclerosis (ALS) is an incurable and progressive neurodegenerative disease impacting motor neurons, leading to bulbar dysfunction in over 80% of cases. Survival rates after diagnosis vary from 20 to 48 months, with only 5 to 10% of patients living beyond 10 years, underscoring the critical need for early detection. However, the absence of reliable clinicopathological markers results in misdiagnoses and up to 12-month delays. Recent studies emphasize speech production reduction as an early indicator of ALS-associated bulbar motor involvement, with 80% to 95% of cases exhibiting speech symptoms like slowness, weakness, and muscle coordination issues. Addressing these challenges, this project proposes an innovative approach utilizing Graph Neural Networks for ALS identification, preserving temporal dependence in speech and incorporating explainable techniques for model interpretation. Additionally, the project involves collaborative research with the Speech-Language Pathology group at the University of Toronto, specializing in ALS detection, aiming to enhance the model through a multimodal strategy involving both audio and video during the overseas research period.