Development of an Edge Computing System for Estimating Root Mean Square and Median Frequency of Surface Electromyogram With Mechanomyograms

Following the development of smart wearable devices and edge computing technologies, a system capable of assessing the condition of the muscle activity may have high potential for application in sport’s fields. Usually, both the root mean square (RMS) and median frequency (MDF) of surface electromyogram (sEMG) have been used to evaluate the isotonic and isometric muscle conditions. However, there exists a major disadvantage in the measurement of sEMG, since it is easily interfered by skin and electrode’s motion during the isotonic contraction. Such interference may decrease the signal to noise ratio (S/N) and the precision of RMS and MDF in sEMG. Thus, the aim of this study is to apply mechanomyograms (MMGs) to estimate the RMS and MDF of sEMG with the machine learning (ML) models. The proposed method was implemented in a STM32 microcontroller (MCU) to achieve a real-time portable system for monitoring muscle activities. The experiment recruited 20 healthy subjects who were requested to walk 150 meters, then to climb seven floors and return, followed by walking another 150 meters. During this process, sensors of sEMG and MMG were attached to the vastus lateralis and gastrocnemius muscles for synchronous signal acquisition. To predict two indexes, RMS and MDF of sEMG, this study extracted five parameters from MMGs, including RMS, MDF, standard deviation (SD), skewness, and kurtosis. In the PC-based analysis, these five parameters were used to train and test the ML models. But, in the STM32-based analysis, the MDF of MMGs was not included due to computational resource limitations. This research compared the performance of several regression models, including Random Forest (RF), CatBoost, XGBoost, Support Vector Machine (SVM), and k-Nearest Neighbors (KNN). The hyperparameter tuning was made via the Optuna tool. The Pearson correlation coefficient (PCC) was used as the primary evaluation metric. The PC-based system without any model limitation showed that the XGBoost model demonstrated the best results with PCCs of 0.911 and 0.902 for estimating RMS and MDF of sEMG, respectively. Its size was 1947 kB and 2085 kB, respectively. The STM32-based system simulated by the PC-based system only using four time-domain parameters and under 1MB of model limitation showed that the XGBoost model provided PCCs of 0.876 and 0.858 for RMS and MDF estimations, respectively. Its size was only 435 kB and 483 kB, respectively. Then, we embedded this XGBoost model in a STM32F756ZG NUCLEO development board for edging computing. Its performance showed PCCs of 0.856 and 0.864 for RMS and MDF estimations, respectively. This study successfully explored an efficient model for estimating RMS and MDF of sEMGs using five parameters from MMGs, and XGBoost with best performance were embedded in STM32F756ZG microcontroller to execute an edge computing for monitoring the muscle activities in real time. PCC and MAPE of XGBoost approached 0.856 and 21.69% for RMS estimation, and 0.864 and 20.04% for MDF estimation. The size of the STM32-based XGBoost for MDF estimation was 483 KB, and for RMS estimation was 435 KB. Therefore, the proposed edge computing models effectively balance the accuracy and resource limitations, which supports a new application in the future portable physiological monitoring system.