STEM, in my opinions is what really sets MAMS apart from most schools. Here we spend six months working on a project to be entered in Febuary fair and hopefully later to WRSEF, MSEF, and ISEF. Each student spends quite a weeks at the start doing basic research to find a topic of their intrest which they will then spend the next few months conducting experiments for.
My project is about maping finger motions in myoelectric prosthetics. Essentially, most prosthetics only allow the patient control of their full hand, not each individual finger. Prosthetics which do allow for this kind of motion however are wildly expensive and require up to 70 surface EMG eletrodes, or require electrodes to be surgically implanted in the patient. My project aims to fix this by using low density EMGs to map individual fingers, proving that this technology does not need to be increadible expensive and should be researched further to make it accesable to all amputees with the required residual muscles.
Electromyographic (EMG) technology is commonly used to observe contractions in muscles throughout the body. This is especially useful for decoding the intended movement for amputees to map the movement to myoelectric prosthetics. However, fine motor movement is very difficult to interpret, especially while using surface EMGs (sEMG) that are not of a high density. This paper proposes a method to decode and map low density sEMG signals taken from the forearm to the intended fine motor movement in the fingers, specifically their contraction and extension. This study used 5 abled bodied subjects and had them perform fifteen different actions while six channels of sEMG data were collected at 250hz. The sEMG data was then used to train a convolutional neural network (CNN), only using data from each individual subject, thus creating a specialized, trained cCNN for each subject. The data shows that the low density sEMG signals processed with a cCNN cannot accurately map fine digit movement to a myoelectric device.
In the United States, over two million people currently experience a lack of limb due to amputations – result of loss of limb after birth- or amelia - being born without one or more limbs. This number increases by 185,000 per year on average (Zhu et. al, 2022). Many amputees choose to purchase prosthetics to offset the related disability with the loss of a limb. However, despite great technological advancements over the past decade, most prosthetic hands do not allow the amputee direct control of each of their fingers individually, and if they do, they rely on embedded electromyography (EMG) sensors or high-density surface electromyography sensors (HDsEMG), which can be very expensive or involve added procedures.
Living without one or more limbs can present a serious disability, especially to those without an upper limb. These transradial amputees, amputees where the amputation occurred below the elbow, loose an incredible amount of dexterity, having one less hand with which to manipulate objects, thus giving their life serious limitations. To offset this disability, many patients opt to get a prosthetic limb, an engineered replacement of the lost limb which is meant to either appear like the lost limb and/or replace its function. The two most common types of prosthetics to purchase are body powered and myoelectric prosthetics. Body powered prosthetics are significantly cheaper as they require no external power but are also very limited in their function, often to only one degree of freedom (DOF). Myoelectric prosthetics on the other hand help offset some of these issues. They use EMG sensors to detect muscle contraction and map it to the intended movement of the subject. The subject’s prosthetic will then reflect this map, thus acting as a fully controlled arm replacement. While these prosthetics have made incredible advancements over the past decade, there are still some issues. Many of these prosthetics can only control up to 2-DOFs when using direct control, the most common commercial control algorithm, and can only control more when using a switching algorithm. Switching algorithms require the patient to contract a separate muscle to indicate to the prosthetic that they wish to change with DOF they are manipulating. This leads to more DOF control but little simultaneous control. AI has been implemented in the past 5 years to help fix some of these issues, both allowing higher DOF simultaneous control and adaptive switching, which predicts which DOF the user wished to use.
Despite these incredible advancements, there remains one glaring issue. Most prosthetics treat the hand as one unit, only allowing it to open and close. This gives amputees back some grasp control, but it remains very fixed. The amputees also cannot manipulate fingers independently, which is used in daily life as a form of communication, such as through pointing. Even prosthetics that do allow for this finger control either require EMGs to be surgically placed into the patient's muscles for more accurate readings or many, over seventy, sEMGs to be connected to their residual stump, thus creating a HSsEMG setup. Both of these options also require a significant residual stump. Thus, my project aims to use a lower density of sEMGs alongside machine learning (ML) to map individual dexterous finger movements to amputee’s prosthetics.
To properly collect and process the data needed to complete the study, many things were needed. First, to collect the data, this study needed a board capable of collecting EMG data through electrodes. As this project was completed under the tutelage of Dr. Kevin Crowthers at the Massachusetts Academy of Math and Science, it had access to the schools openBCI cyton board with a daisy chain, allowing the board to be capable of collecting up to sixteen EMG channels. The Massachusetts Academy of Math and Science also allowed the use of ten electrodes for the purposes of the study. Python 12 was used to program the board along with the brainflow, the API supported by openBCI. The data was then placed into numpy arrays to be fed into a machine learning model. The model was created using pytorch due to its simple nature to program.
There are three programs that will be run throughout this study. The first is a data collection program. This program connects to the cyton board to collect data from electrodes. Throughout this process it collects data into fifteen numpy arrays, one for each position. These arrays will represent 10 seconds of data each. The data is then sent through a bandpass filter at 450hz and a highpass filter at 20hz to remove background frequency noise in the EMGs. After this the data moves to the second program. This program is responsible for processing data, creating the machine learning model, and training the model. The data is split into 1/10 second sets and labels assigned to later be fed into the machine learning training later.
Participants will begin by receiving a thorough description of the study, how their data will be collected, and how their data will be securely stored. Participants will then sign an informed consent sheet, which they will be free to withdraw at any time. In the case of a minor, the informed consent sheet will be sent out prior to the test along with details of the study for their parent/guardian to sign.
To participate in the study, participants will be required to have the forearm of their dominant hand hairless. Having hair in the area of testing can lead to data inconsistent with the intention. If subjects prefer not to remove their hair themselves, a wax kit will be provided to them free of charge at the start of the study. The subject will then be instructed to wash their forearm to prepare for the sEMG placement. The experimenter will then have the subject perform a variety of motions to find the flexor digitorum superficialis, flexor digitorum superficialis and, extensor digitorum communis. These muscles are used for finger contraction and extension and thus what needs to be measured. The experimenter will then place electrode pairs onto the skin of the subject. Two electrode pairs will be placed on the flexor digitorum superficialis, one on the flexor pollicis longus, and two on the extensor digitorum communis The subject will then be shown pictures of the positions they will have to assume during testing, shown in appendix A, and will also be instructed not to move their wrist throughout the procedure. The program will then begin to collect data from the EMG electrodes as the subject moves through the positions. The subject will remain in each position for ten seconds before being prompted to switch. An image of each position will also be provided as that position is assumed.
After running the base model, which features two convolutional layers and four feedforward layers, an average MSE squared loss of 0.207 after ten training epochs. Further results are not yet prepared however if you return tomorrow a graph depicting the MSE squared values and R2 values will be available.
The data collected shows that the engineering goal of this project was not achieved. This result was determined after finding an R2 value of ____ on one of the AI models after sending testing data through. This R2 value shows that the model cannot find significant correlation between the input and output values. After finding this result R2, a two-way ANOVA test was done to compare the average R2 value with that of other models. The null hypothesis in this test was that all models were equal, and the test was to see if any changes made a significant difference. Should the p- values accept the null hypothesis, it would show that all the models are statistically similar and thus all do not fit the value. However, these p-values are too low to make such a claim then it will be known that at least one model is significantly better and thus deserves further testing. The ANOVA test returned a p-value of ____, showing that the models are statistically similar and thus do not fit the data.
Some limitations that could have led to these results include:
• The electrodes used were not medical grade and could only run at 250hz. This could have limited the amount of data the EMGs picked up and thus fed into the model.
• The person applying the EMGs was not a trained specialist. This could have led to the muscles not being properly located and thus not properly measured.
• The electrode patches used were 3cm in diameter. This limits the amount of space to apply more electrodes.
• The cyton board only came with five electrode pairs. This means that if more electrodes were available, more muscles could have been measured while still being considered low density.
The conclusion drawn from this project is that five electrodes cannot gather significant data from the three muscles mentioned above to map individual finger contractions. Despite these results however, there is merrit for further research in this field to understans exactly how many electrodes are needed for individual finger contraction measurement. Knowing this number would allow for prosthetics to be made cheaper for the disabled people who need them most.
Resnik, L., Huang, H. (Helen), Winslow, A., Crouch, D. L., Zhang, F., & Wolk, N. (2018). Evaluation of EMG pattern recognition for upper limb prosthesis control: A case study in comparison with direct myoelectric control. Journal of NeuroEngineering and Rehabilitation, 15(1). https://doi.org/10.1186/s12984-018-0361-3
Hargrove, L. J., Miller, L. A., Turner, K., & Kuiken, T. A. (2017). Myoelectric pattern recognition outperforms direct control for transhumeral amputees with targeted muscle reinnervation: A randomized clinical trial. Scientific Reports, 7(1). https://doi.org/10.1038/s41598-017-14386-w
Zhu, Z., Li, J., Boyd, W., Martinex-Luna, C., Dai, C., Wang, H., Wang, H., & Huang, X. (2022b). Myoelectric Control Performance of Two Degree of Freedom Hand-Wrist Prosthesis by Able-Bodied and Limb-Absent Subjects. IEEE Xplore, 30, 893–904. https://doi.org/10.1109/TNSRE.2022.3163149
Edwards, A. L., Dawson, M. R., Herbert, J. S., Sherstan, C., Sutton, R. S., Chan, K. M., & Pilarski, P. M. (2016). Application of real-time machine learning to myoelectric prosthesis control: A case series in adaptive switching. Prosthetics and Orthotics International, 40(5), 573–581. https://doi.org/10.1177/0309364615605373
Nielson, M. (n.d.-b). CHAPTER 1Using neural nets to recognize handwritten digits. In Neural Networks and Deep Learning. essay. Retrieved from http://neuralnetworksanddeeplearning.com/chap1.html.
Sharma, S. (2022, November 20). Activation functions in neural networks. Medium. https://towardsdatascience.com/activation-functions-neural-networks-1cbd9f8d91d6
Parajuli, N., Sreenivasan, N., Bifulco, P., Cesarelli, M., Savino, S., Niola, V., Esposito, D., Hamilton, T. J., Naik, G. R., Gunawardana, U., & Gargiulo, G. D. (2019). Real-time EMG based pattern recognition control for hand prostheses: A review on existing methods, challenges and future implementation. Sensors, 19(20). https://doi.org/10.3390/s19204596
Harath, T. (2023, May 15). EMG Forearm Position Prediction with OpenBCI Cyton and IDUN Dryodes using Brainflow and Tensorflow. YouTube. https://www.youtube.com/watch?v=y42TazR7UsU
Body anatomy: Upper extremity muscles: The hand society. Body Anatomy: Upper Extremity Muscles | The Hand Society. (n.d.). https://www.assh.org/handcare/safety/muscles#:~:text=Located%20deep%20in%20the%20forearm,though%2C%20contributing%20to%20its%20independence.