Unsupervised Learning for Automatic Classification of Needle Electromyography Signals

Abstract

Introduction. Needle electromyography (EMG) is a diagnostic tool used to identify and localise neuromuscular disorders, but the current evaluation by a neurologist is subjective and requires years of training. Artificial intelligence can be used to automatically classify needle EMG signals. However, previous studies in this area have been subject to bias and have overestimated their performance. As a result, we aim to develop a clinically applicable method for automatically evaluating needle EMG signals to provide a more objective and efficient means of analysis. Methods. In this study, we implemented and evaluated two unsupervised learning models, a convolutional autoencoder + k-means clustering model and a deep convolutional embedded clustering model. The models were evaluated on two classification tasks: the classification of spontaneous, voluntary and insertional activity and the classification of spontaneous activity in needle EMG data classified as rest. We used hospital-acquired needle EMG data from the Amsterdam University Medical Centre (UMC), location AMC, from a total of 326 patients. The data was converted to Mel spectrograms, resulting in a total of 1.3 million images available for training. Results. The unsupervised learning models reached 93.5% accuracy on unseen test data. The classification of phenomena present in rest, such as fibrillation potentials and positive sharp waves, was less successful and requires further research. Conclusion. Our study provides valuable insights for the use of unsupervised learning in the automatic classification of needle EMG signals and highlights the need for further research in this area.