Classifying muscle states with ultrasonic single element transducer data using machine learning strategies

Being able to distinguish non-invasively between different muscle states is crucial for rehabilitation and sports athletes alike. The analysis of muscle activities is often performed using optical systems, kinetic approaches or surface electromyography. However, these methods can only obtain information from the body surface. In this work, raw ultrasound radio frequency data is used for muscle state classifications as this method provides information from deeper muscle layers. A setup to classify muscle contractions with artificial neural networks and traditional time series analysis algorithms is presented. Experiments with the pulser-receiver Olympus 5800PR and Panametrics transducers were performed to obtain A-scans from the calves of healthy volunteers. Dimensionality reduction techniques, such as (Kernel-) Principal Component Analysis and t-Distributed Stochastic Neighbor Embedding were applied, which provide information about the distribution of A-scans belonging to relaxed and contracted muscle states. A comparison of different ML methods is presented with average F1 scores of up to 89 % after appropriate post-processing for muscle contraction classifications and average F1 scores of up to 76 % for muscle fatigue state classifications. We conclude that low-cost ultrasound measurements can be used for reliable muscle activity tracking. More data is expected to result in even more robust future solutions.