An Artificial Intelligence Approach to Augmentative and Assistive Communication for Patients with Amyotrophic Lateral Sclerosis

Amyotrophic Lateral Sclerosis (ALS) progressively impairs motor functions, making communication increasingly difficult for affected individuals. However, many patients with ALS retain control over their ocular and periocular muscles, providing an unique opportunity for Augmentative and Alternative Communication (AAC) systems. This work introduces a calibration-free surface electromyography (sEMG) decoding module, developed as part of the HALO system, an AAC solution that combines sEMG with Large Language Models (LLMs). The HALO system harnesses the power of the LLMs to generate personalized, context-aware response options. The user can then select a full-sentence reply with a single eyebrow muscle contraction. Eyebrow muscle activity is captured via textile electrodes embedded in a headband and processed in real time using a Recurrent Neural Network (RNN). Trained and evaluated on data from patients with ALS, the model achieved 96% accuracy and an 85% F1-score. While direct comparisons to existing systems are limited --since existing sEMG-based AAC systems are rarely benchmarked on ALS cohorts and fail to report key metrics such as gesture detection accuracy -- our results demonstrate that the models reliable gesture detection can support efficient, low-effort communication for individuals with severe motor impairments.