Subject-Independent Automated Detection of Amyotrophic Lateral Sclerosis from Electromyography Signals Using ALSNet

Abstract

Amyotrophic Lateral Sclerosis (ALS) is a progressive disorder of the nervous system that affects motor neurons and causes progressive weakness and loss of function of motor neurons. Clinically relevant information on neuromuscular activity is provided by Electromyography (EMG) and can be used to assist the diagnosis of ALS. The use of EMG interpretation can be difficult due to their complex temporal and morphological characteristics, however. In this study, the novel approach of a deep learning method (ALSNet) to automatically classify ALS and Normal EMG signals is explored. The experiments were performed with processed EMG data extracted from publicly available N2001 EMGLAB database. All 302 EMGs (151 Normal, 151 ALS) were divided into 3,322 one-second EMG segments. A subject-independent partitioning strategy for the dataset was used. ALSNet is a deep network developed based on Keras and TensorFlow frameworks, and trained with Adam optimizer by using binary cross-entropy loss. The accuracy, sensitivity, specificity, precision, F1-score, and balanced accuracy were used for evaluation of the model. ALSNet obtained 97.74% accuracy, 96.77% sensitivity, 98.59% specificity, 98.36% precision, 97.56% F1-score and 97.68% balanced accuracy on the independent test set. The findings show the promise of deep learning analysis of EMG signals for automated ALS detection and offer a foundation for future analysis with more and varied clinical datasets.