MACHINE LEARNING TECHNIQUES APPLIED IN SURFACE EMG DETECTION-A SYSTEMATIC REVIEW

Surface electromyography (EMG) has emerged as a promising clisnical decision support system, enabling the extraction of muscles' electrical activity through non-invasive devices placed on the body. This study focuses on the application of machine learning (ML) techniques to preprocess and analyze EMG signals for the detection of muscle abnormalities. Notably, state-of-the-art ML algorithms, including Support Vector Machines (SVM), k-Nearest Neighbors (KNN), Artificial Neural Networks (ANN), Random Forests (RF)


INTRODUCTION
Electromyography (EMG) is a technology for analyzing and assessing electric activities/conditions developed by muscles.EMG of muscles is performed by using an electromyograph to provide data called an electromyogram.The device detects electric signals of the muscle's cells during their electrical and neurologically activated motion (Reaz, Hussain, & Mohd-Yasin, 2006).

Figure.1 EMG Working diagram
In figure 1, the working diagram of EMG is provided.Electrodes/Device is connected to the human surface for muscle electric protentional detection.The electric signals in the form of positive and negative signals are pre-processed by a powerful computing device, after that, an amplifier converts the electrical signals, and the results show on the computer monitor.
There are two types of EMG, namely: Surface EMG and Intramuscular EMG (Szyszka-Sommerfeld, Sycińska-Dziarnowska, Spagnuolo, & Woźniak, 2023).The first kind of EMG (surface EMG) places a metal on the skin of the living thing's body and extracts muscles' electrical signals.In simple words surface EMG is working on the surface of the skin.The second kind of EMG is Intramuscular EMG which works inside of the skin and sends electrical signals for further analysis.For surface EMG two electrodes are required to differentiate between voltages.
Surface EMG (sEMG) is a non-invasive technique used to measure the electrical activity of muscles.It has been widely used in the field of physiotherapy for monitoring muscle activation levels during exercises, as well as in sports to improve athletes' performance and reduce the incidence of soft tissue injuries.In recent years, machine learning techniques have been applied to sEMG data to improve the accuracy of muscle activation detection and classification.These techniques use algorithms to automatically detect and classify sEMG signals, making the analysis process faster and more accurate.This is important for identifying the timing of muscle activity during movements and exercises, which can help physiotherapists and coaches to optimize training programs and prevent injuries.By analyzing sEMG signals from different muscles, machine learning algorithms can classify movements and identify specific muscle activation patterns associated with different tasks.The application of machine learning techniques to sEMG data has the potential to revolutionize the field of physiotherapy and sports medicine.By improving the accuracy and speed of muscle activation detection and classification, these techniques can help clinicians and coaches to optimize training programs and prevent injuries, leading to better outcomes for patients and athletes.In this paper, we will focus on the first kind of EMG (Surface EMG) detection using Machine Learning techniques.
S-EMG monitor during intended muscle movement can be as a signal, while the connection of all active motor units (MUs) is called an interference pattern.Interference means that the signals are mixed with different wavelengths and dimensions.The Surface EMG signal characteristics largely depend on pattern and interdependency (Stegeman, Blok, Hermens, & Roeleveld, 2000).Motor Unit (MU) is basically the complete model for S-EMG regarding the interference pattern as a linear summation.
Different kinds of approaches are used but the two are the most popular approaches to studying the connection between the S-EMG and the characteristics of the neuromuscular system (Farina, Merletti, & Enoka, 2004).The first is the conventional method, while the second is the backward method.The initial method was using models to forecast how various physiological processes would influence S-EMG characteristics.In the inverse method, electromyography (EMG) is used to discover the physiology at play.The accuracy of Surface-EMG measurements relies heavily on the converter and amplifier design, as well as the electrodes' interface with the skin, and afterward storing of the signals in digital form from their analog form (A/D Conversion).Measured EMG signals and Noise are two measurements of the electric signals for analysis.The main goal of Surface EMG detection is to increase the signal amplitude and to decrease noise for correct results for the muscles (Day, 2002;Ferrario, Tartaglia, Galletta, Grassi, & Sforza, 2006).Connecting the electrode devices to the skin surface correctly can reduce noise.
Motor Units (MUs) play an important role in surface-EMG, as when the number of Mus increases then the difference between noise and amplitude can be seen by necked eyes (De Luca, Adam, Wotiz, Gilmore, & Nawab, 2006).In the previous decades, the effect of the electrodes on a number of estimation conduction velocity (CV), spectral variables, and amplitude of the Surface-EMG has been addressed in various studies.The considered muscles range from the muscles of the shoulder, arm, and leg to the muscles of masticatory.The most significant effort to standardize was made in 1997-1999 within the European project on "Surface EMG for Non-Innovators".
"Intelligent agents" are defined as "tools that recognize their environment and take acts that maximize the probability of successfully achieving their goals" (Crevier, 1993).The field of Artificial Intelligence (AI) has been around since the latter half of 1956.Enhancing whiz systems was the first step in applying AI in the medical field; from there, rules were gleaned via discussions with doctors and then implemented into software (Zhang, 2020).In 1976, an initial expert system called "MYCIN" was developed to recommend antibiotics for bacterial illnesses; it followed close to 450 rules/SOPs.The sheer number of rules made the expert system impractical for application in actual hospitals.Machine Learning was created (Zhang, 2020) to overcome the constraints of expert systems.When it comes to clinical decision support systems, ML's ability to replace manual rules with system-generated ones is invaluable.This is because ML algorithms may learn from their surroundings and apply what they've learned in the future.Deep learning-based artificial neural networks (Hinton, 2018) are the most well-known and efficient ML approach currently available.Because ML relies so heavily on having access to large amounts of highquality data, it is frequently referred to as Data-Intensive Systems.The proliferation of low-cost, Internet-enabled electronics has led to a surge of healthcare data.Expert systems dominate the clinical decision-support space, although ML-based CDSS are gaining traction thanks to their efficient learning and prediction algorithms (Rawson, Ahmad, Toumazou, Georgiou, & Holmes, 2019).X-rays, MRI scanners, and the identification of malignant nodules in the lungs and other tissues are only some of the many applications of ML in the systems used by medical professionals.In this work, we will examine the effectiveness of Machine Learning strategies for detecting Surface-EMG.
Surface EMG detection using Machine learning is the recent advancement in EMG patterns evaluation for accurate results (Phinyomark & Scheme, 2018).In Machine Learning we can train datasets and acquire accurate results from raw data, and it has three steps namely: A Model, A Dataset, and the Hardware (Ahamed, Benson, Clermont, Osis, & Ferber, 2017;Ahamed et al., 2018).In recent times, Machine Learning Algorithms are applying to analyze the skill of myoelectric prostheses or any other rehabilitation device.As a result, the characterization of two prime muscles is difficult to diagnose neuromuscular diseases (Islam et al., 2014).Machine learning techniques will be best to accurately analyze the above diseases of neuromuscular (Gu, Yang, Huang, Yang, & Liu, 2018).There are different Machine Learning algorithms that are used to investigate muscles disorder using Surface-EMG signals with a single ML technique.A Support Vector Machine (SVM) classifier was commonly used for EMG-based muscles pattern analysis recognition (Saponas, Tan, Morris, & Balakrishnan, 2008;Yoo, Park, & Lee, 2018).In (Benalcázar, Jaramillo, Zea, Páez, & Andaluz, 2017) the author proposed real-time muscles recognition system using five types of hand gestures (Fist, Open, Wave-In, Pinch, and Wave-Out) using Surface-EMG while sensors were plugged-in the forearm and used the ML (K nearest Neighbor Classifier) technique.Some other researchers used ML algorithms as a single classifier such as Artificial Neural Network (ANN) (Su, Chen, Cao, & Zhang, 2016;Uvanesh et al., 2016), Random Forest, and Bayesian Classifier (Chen et al., 2007).Upper limb muscles are also analyzed by different researchers by using machine learning algorithms (Kim, Choi, Moon, & Mun, 2011;Pancholi & Joshi, 2018).In this paper, we provided a detailed overview of the available work on Surface-EMG detection using machine learning algorithms.As in the available literature, no one provided a detailed survey paper on Surface-EMG detection using Machine Learning techniques with the latest issues and future directions.
The outcomes of the proposed study are as follows: • We discussed the various Machine Learning strategies that may be used for Surface-EMG detection.
• Provided the latest issues in the field of Surface-EMG detection and future directions.Further, the paper is planned as follows: In Section 2 we provided Method, in which we detail explained Article Search Procedure, Article Inclusion, and Exclusion Criteria, Data Extraction, Article Search Results (PRISMA flowchart).In Section 3 we provided detailed literature on the selected papers.Section 4 consists of the discussion and future directions.Section 5 consists of the conclusion of the systematic review paper.

METHOD
In this section, we provided a detailed explanation of Article Search Procedure, Article Inclusion, and Exclusion Criteria, Data Extraction, and Article Search Results (PRISMA flowchart).In the following sub sections, we provided the method of the systematic study on Surface-EMG detection using Machine Learning Techniques: Article Search Procedure: In this sub-section, we provided an article search procedure.We used a systematic searching procedure to find out related papers in the field of Surface-EMG detection using Machine Learning Techniques.In the searching process, we targeted different databases such as, (Scopes, Web of Science, PubMed, Science Direct, and IEEE Xplore) using three keywords "Surface EMG", "Machine Learning", and "electromyography (EMG)".We have selected conferences and journal papers, and we also selected some book chapters.Article Eligibility Requirements: In the searching procedure, we found different papers, but we used inclusion criteria based on the two words namely: Surface EMG, and Machine Learning.We studied the papers thoroughly such as their title, abstract, proposed model, and conclusion.When we found the papers mostly related to our title then we added them to our literature and introduction.Such articles were excluded that were related to EMG or Intramuscular EMG the second kind of EMG (Szyszka-Sommerfeld et al., 2023).Articles were selected that were related to the first kind of EMG (Surface EMG).Data Extraction: The selected papers were then read carefully and summarized the important information about their used machine learning techniques and proposed method with results.The main criteria for studying papers were the following: In the flowchart, we provided the different databases that are searched for Surface EMG papers selection using Machine Learning techniques.Furthermore, a total of 676 papers were found in the databases using the keywords "Surface EMG", "Machine Learning", and "electromyography (EMG)".After applying a filter of conferences papers, journal papers, and book chapters, the articles were reduced to 531.Similarly, exclusion and inclusion criteria are applied such as to remove duplicate papers, not related to the topic (such as Machine Learning), and not related to the first kind of EMG (Surface EMG) then 47 papers remained.The further filter is used for a detailed study of the remaining papers and removed 16 papers, then we have the remaining 31 papers, which we discussed in detail in section 3 of the literature of Machine Learning Techniques used for Surface EMG detection.In figure 4 we also provided keywords used for searching and years as well.Nowadays it is basically used for clinical decision support systems, especially for surface EMG detection and giving better results.In (Liarokapis et al., 2013) the author used ANN as a hidden layer for surface EMG detection classification and compared it with the other machine learning techniques such as RF and SVM.In the environment where fewer data were available, the ANN outperformed while where huge data were available then SVM and RF performed well.

Figure. 7 ANN general architecture
In (Tibold & Fuglevand, 2015) the author used ANN to predicate Surface EMG signals on muscles using loaded and unloaded situations.They used four layers of ANN for classifications.From the result, they concluded that ANN is the best machine learning technique where when we received complex patterns of muscles.Similarly, another author (Ahsan, Ibrahimy, & Khalifa, 2011) used ANN for backpropagation and their success rate was 88.4% to recognize the complex patterns for analyzing EMG signals.Moreover, in (Jose, Raj, Adithya, & Sivanadan, 2017) the author compared and presented the accuracy rate for complex patterns in muscles detection for ANN was 91.6 % while for Random Forest it was 97.7%.In (Saeed et al., 2021)   In this survey paper, several papers are selected in which KNN was the primary ML technique used for surface EMG detection.Such as, in (Kim et al., 2011) the author proposed a comparison-based study in which they used Quadratic discriminant analysis, linear discriminant analysis, and KNN.They used the above techniques for wrist motions detection using surface EMG.In the proposed study the author concluded that KNN performs well in comparison with the other two classifiers.Similarly in another study (Paul et al., 2017), the author provided that SVM ML techniques beat KNN in terms of accuracy of surface EMG detection.Moreover, in (He, Yang, Wang, Cheng, & Hu, 2017) the author proposed a new method for analyzing hand gesture based KNN technology.They classify surface EMG signals and provide more accuracy as compared with simple KNN-based technology.From the study, it is concluded that in upper limb muscles KNN performed very well in terms of accuracy.In (Tuncer, Dogan, & Subasi, 2020) the authors used a novel ternary pattern and discrete wavelet (TP-DWT) method for surface EMG signals detection.They used three levels such as low, high, and moderate.The accuracy rate they achieved were 97.78 %, 93.33 %, and 92.96 % respectively.Similarly, in (Khan, Alam, Khan, & Farooq, 2021) the author used the KNN technique for features extraction and found that regression coefficient and Willison Amplitude features performed well.
Naïve Bayes (NB): Naïve Bayes classification is a classification-based algorithm used for features classification.It is based on the Bayes theorem and used in the field of medicine for clinical decision support systems and other parts of medical fields.In (Wahid, Tafreshi, Al-Sowaidi, & Langari, 2018b) the author compared SVM, KNN, RF, DA, and NB, from the comparison result the author proved that NB performed well where the dataset was defined proper, while in all other situations SVM and RF performed well.In the following table 1 we have provided the detail summary of the discussed Machine Learning Model Tunable Q wavelet transform (TQWT) Boosting of Support Vector Machine (SVM) 100.00 The following challenges are identified through the study of the papers: • Machine Learning based algorithms are required to identify the damage area and to detect early Alzheimer diseases.
• Machine Learning based techniques and studies are required to forecast and train the already data.• Hybrid classification models are required to increase the detection accuracy.The following table 2 is used for the abbreviations of the terms used in this paper: The amalgamation of fuzzy logic with ML techniques holds the potential to yield improved accuracy in the detection of muscular activities.Moreover, the exploration of cloud and fog computing as means for dataset storage and computation offers promising avenues to enrich the available datasets dedicated to surface EMG detection.By harnessing the capabilities of cloud and fog computing, researchers can access diverse and substantial datasets, consequently contributing to the advancement of surface EMG detection methods and fostering progress in the field of biomedical research.

CONCLUSION
In conclusion, the utilization of surface electromyography (sEMG) in the study of muscle activation patterns has provided valuable insights into the modular organization of multi-muscle actions across diverse motor tasks.Through the identification of fundamental synergies, characterized by simple patterns with varying individual weights, the nervous system appears to streamline muscle regulation to achieve a versatile and efficient control strategy.This finding holds substantial implications for the fields of motor control and neurorehabilitation, suggesting that the central nervous system optimizes force and motion generation by leveraging these foundational synergistic patterns.
Despite the significant potential of sEMG-based investigations, estimating force sharing among synergic muscles remains a challenging endeavor.The fluctuations in individual muscle forces, as captured by sEMG amplitudes, complicate the accurate determination of their precise contributions to net joint torque.Over the past three decades, numerous efforts have been made to elucidate force distribution, yet the complexity of neuromuscular coordination during dynamic tasks has hindered precise estimations using sEMG alone.
To address this challenge, future research endeavors may benefit from complementary approaches that integrate sEMG data with advanced biomechanical analyses or leverage cutting-edge machine learning techniques.By combining multiple data modalities, a more comprehensive understanding of force sharing among synergistic muscles can be achieved, unlocking new avenues for the optimization of rehabilitation protocols, enhanced athletic performance, and the advancement of bio-inspired robotics.
Figure. 2 No of papers according to the search engines Figure. 4 Prisma Flowchart of the systematic searching procedure of the paper's selection Literature based on Surface-EMG Detection Using Machine Learning: In this section, we provided a detailed overview of the literature on ML Techniques used for Surface EMG detection in the medical field.The following ML techniques are used to detect Surface EMG: Support Vector Machine (SVM): SVM is an old machine learning technique used for classification and provided better results as compared with other available machine learning techniques.Many authors used the SVM technique for surface EMG signals detection and found a very accurate and efficient technique.In our systematic review on surface EMG, we found several SVM-based techniques used which we provided in detail.In (Kakoty & Hazarika, 2011) the author used the kernel-based Support Vector Machine technique and normalized the surface EMG signals into six different clenches.From the results, the author concluded that they achieved 97.5% accuracy through fold cross-validation.Similarly, in (Pomboza-Junez & Terriza, 2016) the author used SVM for surface EMG signals for upper limb

Figure. 6
Figure.6 Random Forest general architecture the author proposed an ANN-based model, which combined ANN and LDA for the classification of EMG signals.Their accuracy was 99.27% & 93.64% respectively for surface EMG.The results showed that combined EMG and S EMG signals performed with better accuracy up to 98.24% and 89.25%.Similarly, in (Li et al., 2023) the author used GRNN neural network for surface EMG signals detection.They used the proposed model for redundant information reduction of the signals.The accuracy rate was noted as 95.1% for EMG signals.In (Qi, Jiang, Li, Sun, & Tao, 2020) armband and sliding window, the concept was used for surface EMG signals detection, to extract the features.They used a real-world dataset and the accuracy rate provided was 98.7%.

Figure. 8
Figure. 8 KNN general architecture • Different Machine Learning Techniques such as KNN, ANN, RF, and DCN, BC ET are discussed in detail applied in the field of Surface EMG detection used for muscles pattern recognition.• Detail overview/literature is provided on Surface EMG using Machine Learning Techniques with pros and cons.• At the end of the paper, we summarized the results of the available work and discussed them in detail with comparison.

Table 1 .
Summary of the discussed algorithms