Comparing Kolmogorov-Arnold Networks and Conventional Machine Learning Algorithms for sEMG Based Movement Classification

Abstract

Myoelectric prosthetics utilize machine learning(ML) algorithms to translate surface electromyography(sEMG) signal patterns into movement, allowing patients to easily control their prosthesis. While the use of conventional ML algorithms has been effective within these prosthetics, challenges still remain in terms of classification accuracy and computation speed. It was hypothesized that Kolmogorov-Arnold Networks (KANs), a novel neural network known for learnable activation functions, could be used to address these issues.

To test this hypothesis popular KAN libraries were evaluated for their ability to accurately classify sEMG signals from the NinaPro DB2 dataset. Various model parameters and architectures were tested for their ability to produce accurate and efficient movement classification results. Over the course of KAN testing the efficient-kan model was identified as the most successful KAN classifier. Efficient-kan was then tested against conventional ML algorithms such as MLP, KNN, RF and others.

Results revealed that the efficient-kan model achieved classification accuracy comparable to several conventional ML algorithms. However, efficient-kan did not outperform the top performing conventional classifiers in terms of classification accuracy or computational efficiency. These project findings indicate a need for the further development of KAN algorithms before they are able to address current issues within myoelectric prosthetics and be considered a viable replacement for established ML algorithms.

Popular Abstract

Are Kolmogorov–Arnold Networks the key to achieving more accurate sEMG classification?

For years myoelectric prosthetics have allowed users to regain a limb. However this new limb comes with a list of accuracy related warnings and restrictions. Don’t operate motor vehicles, don’t hold hot drinks, and more. This list needs to be shortened to allow users to fully integrate with their prosthetic. Could Kolmogorov-Arnold Networks(KANs) be the solution to this problem?
Answer: not at the moment.

Within myoelectric prosthetics, electrodes in the prosthesis record surface electromyography(sEMG) signals. These signals are then translated by machine learning(ML) algorithms into movements performed by robotic components. Conventional ML algorithms have worked well in this field but their accuracy still leaves the prosthetic user restricted by device warnings. KANs are a new algorithm that have shown promise within complex classification problems. KANs are unique in that the activation function that they apply to input data to produce classification predictions is not constant but is a learnable function. As the algorithm is trained the function adapts to reflect the complex patterns connecting input data to output data.

Five popular KAN variants were tested for their ability to classify sEMG data. Pykan, the original KAN! Pykan-sparse, the original with a twist, a strong mask. Efficient-kan, a faster version of the original. Efficient-kan_DROP, the fast version with the ability to generalize. And lastly ChebyKAN, a unique KAN applying a slightly different technique to calculate activation function. After various algorithm tweaks and tests classifying sEMG data from the NinaPro DB2 database, it could be concluded that efficient-kan and efficient-kan_DROP most often correctly
classified movements. With efficient-kan producing an accuracy of 73.7%(average across three exercises) and efficient-kan_DROP an accuracy of 75.5%(average across three exercises). In regards to the very complex data set, these prediction accuracies are quite high!

After this result, efficient-kan was directly compared to conventional ML algorithms such as multilayer perceptions(MLPs), k-nearest neighbor(KNN), and five more. Again the algorithms were tested on various sets of sEMG data. Even though we had high hopes for the efficient-kan classifier it did not outperform all conventional classifiers. It was outperformed by the MLP classifier by approximately 3%. However, the algorithm’s accuracy exceeded simpler ML algorithms and performed on par with other conventional classifiers.

This outcome led to the conclusion that KANs currently don’t have the performance accuracy to replace conventional algorithms in myoelectric prosthetics. Only the most accurate algorithms should be used in the prosthesis. However, as the classifier is still new, there is room for improvement and further testing. Work done within this project will hopefully be used as a basis for the further testing that shortens the list of prosthetic warnings and restrictions.

• author: Johnson, Mia Britta
• supervisor: Christian Antfolk
• organization: Department of Biomedical Engineering
• course: BMEM01 20251
• year: 2025
• type: H2 - Master's Degree (Two Years)
• subject: Technology and Engineering
• keywords: Kolmogorov-Arnold Networks, sEMG, myoelectric prosthetics
• language: English
• additional info: 2025-14
• id: 9206210
• date added to LUP: 2025-06-30 12:55:34
• date last changed: 2025-06-30 12:55:34