Introduction: Electromyography generates an electrical current produced from neuromuscular junctions during muscular movement, but signals are inherently noisy as a result of moving through tissue [1]. Neurodegenerative diseases like ALS, MS, Parkinson’s, and intrinsic tremors result in varying levels of motor neuron degeneration leading to tremors and potential loss of muscular function [2]. However, EMG signals are still present in areas with affected motor neurons, and rehabilitation research is a large area of interest for improving conditions for those affected by such diseases. Non-invasive methods can record surface EMG (sEMG) signals for processing and rehabilitation. Most prosthetic approaches seek to either stabilize or assist with load bearing. This work sets out to create a low-cost fixed robotic arm attachment with algorithms to stabilize and reduce load on the user as a precursor to an assistive wearable that replicates bicep motion. The device mimics the user’s elbow flexural motion by converting sEMG signals into an arm angle using two contrasting computational approaches: a Deep Neural Network (DNN) and traditional signal processing algorithms. The prototype had three main areas of focus. The first served to test design and usability of a potential prosthetic with form factor constraints for weight, force capability, durability, and cost. Next, the focus was quantitatively on precision and resolution of arm angle. Finally, a working prototype focused on qualitative comfort and device ease of use.
Materials and
Methods: To create the physical prototype, two high torque servos were used to actuate the 3D printed arm along with a circuit for analyzing sEMG signals. Signals were amplified and filtered through surface electrodes from a passband of 5-500Hz and a notch filter at 60Hz. The signals were then sampled via a NRF52833 chip microcontroller ADC at 2000Hz. Once signals were acquired, several datasets of users moving their arms in arbitrary motions were recorded. sEMG signals from the circuit and the ground truth angles of the arm were manually labeled. Using this data, several algorithms were tested for smoothing and calculating arm angle output from sEMG. A running average of the peak-to-peak EMG signal performed the best in practice for predicted arm angle output compared to a pure averaging and a RMS algorithm. In addition, to avoid motion artifacts, the algorithm bins the calculated angle by rounding the predicted angle to the nearest 8 degrees. The microcontroller algorithm solely required a quick reading from a flexed and unflexed state to calibrate. To compare results from the microcontroller algorithm and provide a frame of reference, the 3-Layer DNN was constructed using supervised-learning to highlight a more robust and computationally expensive alternative. The DNN was trained on labeled data of 0, 45, 90, 135, and 180 degrees across 30,000 data points. The training and data collection took around 10 minutes total and must be calibrated per individual by recording data at the 5 aforementioned angles.
Results, Conclusions, and Discussions: After constructing the prototype, we ran experimental trials of arm angle vs. peak-to-peak voltage resulting in a calibration curve with |r| = 0.99. This allowed for a linear approximation of arm position with respect to sEMG signal in the microcontroller algorithm. The algorithm had an R^2=0.75 in comparison to the R^2=0.85 for the DNN. In addition, the myoelectric arm peak-to-peak algorithm was subject to noisy motion artifacts and jitter. To control the noise, we restricted the arm to 8 degree bins that minimized jitter at the cost of arm angle resolution. The neural network can calculate arm angle to the nearest degree, giving it an advantage in resolution. Although the myoelectric arm had a lower accuracy and resolution in mimicking the ground truth compared to the neural network, it has tradeoffs in performance. Specifically, the prototype algorithm can be calibrated in about 1 second with two button presses while the neural network algorithm was recorded taking 10 minutes to collect data and train. The neural network also needs a large amount of data for training. Both algorithms can run in real-time with the neural network computing inference and the microcontroller algorithm computing a running average. There is an inverse relationship between the window size duration for averaging and average error with the microcontroller, but a window size of 400 ms was chosen as it appears real-time while still providing the R^2 of 0.75. In conclusion, while the neural network takes longer to train and calibrate, it provides the most accurate arm angle prediction results. Future work for the project involves embedding the neural network onto a high-computation microcontroller. The network training and calibration can eventually be parallelized and automated to a few minutes per user. One major drawback is to address signal acquisition. Analog circuitry and filters provide a clean signal, but multi-layered PCBs are needed to isolate power and signal planes. Iterating on all these features will make the device fully wearable once mechanical components are adjusted to fit the user.
Acknowledgements (Optional): References [1] Raez, M. B., Hussain, M. S., & Mohd-Yasin, F. (2006). Techniques of EMG signal analysis: detection, processing, classification and applications. Biological procedures online, 8, 11–35. https://doi.org/10.1251/bpo115 [2] Khodadadi, V., Rahatabad, F. N., Sheikhani, A., & Dabanloo, N. J. (2023). Prediction of Biceps Muscle Electromyogram Signal Using a NARX Neural Network. Journal of medical signals and sensors, 13(1), 29–39. https://doi.org/10.4103/jmss.jmss_3_22
