L-SHADE optimized learning framework for sEMG hand gesture recognition

The hand plays a vital role in human interaction with the external environment. A hand can convey the emotions of a human through different hand gestures and can sensitively perceive changes in the environment. The hand affects the daily life activities of amputees. Amputations mainly result from pathology, injuries, and accidents^1,2. As per the World Health Organisation (WHO), estimates that approximately 16% of the global population, or 1.3 billion people, live with a significant disability³. Of these, arm amputation affects about 3 million people globally and accounts for almost 2.4 million cases recorded in developing countries. In recent years, engineering applications in the biomedical and robotic fields have become more prevalent, mainly in medicine, remote surgery, assistive technologies, pacemakers⁴, etc. Due to the growing population and demand, and for some specific applications like assistance to physically impaired people or remote surgery, the assistive tools need to be highly precise and accurate. For instance, writing on paper with a prosthetic limb or performing remote surgery requires greater flexibility and accuracy. Therefore, a robotic hand is among the most effective solutions for armless and handless people⁵. To create an autonomous system for amputees, precise control of the robotic hand using signals from the human body or consciousness is required. Researchers have investigated the use of electromyogram (EMG) and electroencephalogram (EEG) signals in this scenario⁶. The EEG records the brain’s electrical activity and is captured by placing electrodes directly in contact with the human head. The EMG, on the other hand, records electrical signals generated by skeletal muscles⁷.

EMG is the signal utilised most frequently in prosthetic applications⁸. Using EMG to differentiate between hand gestures is more straightforward than other signals due to the physiological processes underlying its development in the skeletal muscles. As a result, biological applications and clinical diagnostics are the primary motivations for using EMG signals. There are two methods for collecting the EMG signals from muscles: invasive and non-invasive. In invasive methods, needles or wires are injected into the muscles to collect the muscle action potential, whereas, in the case of non-invasive methods, electrodes are pasted on the skin surface, from where the field generated by muscle fiber potential is collected⁹. The signal collected from the invasive method is called intramuscular EMG (iEMG), and the non-invasive method is called the surface EMG (sEMG) signal. The sEMG signal is frequently employed because iEMG may sometimes rupture tissue and cause discomfort^10,11.

Adopting assistive technology based on sEMG signals can improve the quality of an amputee’s day-to-day life. Pradhan et al.¹² have presented the use of INA128 to collect sEMG data for capturing hand gestures using Texas ADS1294, which is specifically designed for biomedical applications. Pancholi et al.¹³ utilized the ADS1298 IC, similar to the ADS1294, for biomedical signal acquisition. Ankit et al.¹⁴ considered the Myoware sensors to collect sEMG data for the activity classifications. Even though there are several signal conditioning tools and high-level EMG signal collection devices on the market, such as Biopac and BioNomadix, they are known for making high-quality, research-grade EMG devices and have a long-standing reputation for accurate signal capture and reliable data¹⁵.

The EMG signals collected through these acquisition devices for different activities, such as hand gestures, gait activities, etc., can be used for real-time applications such as robotics and prosthesis control¹⁶. Ulkir et al.¹⁷ demonstrated a fuzzy logic-based classification of acquired muscular signals, which is easy while dealing with three handcrafted features (Root mean square, Wavelength, Kurtosis) to classify signals. However, new fuzzification rules must be prepared as the number of features increases, and the task becomes complicated. To overcome these, machine learning is introduced as a solution. Maria et al.¹⁸ have collected the data with the help of four-channel acquisition devices for six different hand gesture activities. They extracted six time and frequency domain features, such as Willison Amplitude (WAMP), Variance (VAR), Mean Absolute Value (MAV), Waveform Length (WL), Median Frequency (MDF), Mean Frequency (MNF), and then applied three machine learning models for the classification of different gestures. Kuzborskij et al.¹⁹ consider the publicly available sEMG signals of 52 different gestures performed by twenty-seven participants and extracted seven time and frequency domain features. The gestures are then classified and identified using four distinct machine learning classifiers, namely Linear Discriminant Analysis (LDA), Multilayer Perceptron (MLP), k-Nearest Neighbour (KNN), and Support Vector Machine (SVM). Omari et al.²⁰ extracted ten handcrafted features and analyzed the impact of the combinations of features with four machine learning models to evaluate the best combinations of features with the highest classification rate for gesture recognition. Song et al.²¹ used a wearable smart sEMG recorder integrated with gradient boosting to recognize the hand gesture.

According to the literature, machine learning models utilize hyperparameters, which have preset parameters. These parameters may be configured during the training phase of the model as well as the initial configuration phase. Due to the default value of these models, it does not guarantee the maximum level of efficacy²². As a result, machine learning models with an extensive range of hyperparameter values can be considered to construct the best-performing machine learning model. Determining the optimal hyperparameter value for a machine learning model is often referred to as hyperparameter tuning²³. Sajjad Nematzadeh et al.²⁴ have utilized Genetic Algorithm (GA) and Grey Wolf Optimization (GWO) based hyperparameter tuning of various machine learning algorithms. In this classification, diverse types of biological, biomedical, and natural data sets are classified, including molecular interactions, clinical diagnosis, cancer, RGB images of human skin, behavior-related predictions, and X-rays of COVID-19 and cardiomegaly patients. B. Chitra and colleagues²⁵ use a densenet121 deep learning model with hyperparameter optimization using Atom Search Optimization (ASO) to detect cervix cancer. Marius Geitle et al.²⁶ have conducted a comparative analysis of three optimization techniques (L-SHADE, random search, and adaptive random search) with an XGboost machine learning model to modify the hyperparameter and found that L-SHADE outperforms the others for different publicly available datasets. K. Jayaprakash et al.²⁷ have implemented the Artificial Rabbits Optimization (ARO) with deep learning models for plant disease classification. Woo-Young Lee et al.²⁸ have used the Harmony Search (HS) to tune the hyperparameters of Convolutional Neural Network (CNN) for image classification. Nazrul Islam et al.²⁹ have presented the water quality prediction based on Elman Neural Network (ENN) and tuned the hyperparameters using Artificial Ecosystem Optimization (AEO). Ismail Damilola Raji et al.³⁰ have applied Stochastic Gradient Descent (SGD), Bayesian Optimization (BO), GA, Particle Swarm Optimization (PSO), and Biogeography-Based Optimization (BBO) to tune the hyperparameters of different machine learning models for the benchmark UCI and MNIST datasets. Xiuwu Sui et al.³¹ have classified the six upper limb activities using PSO with improved SVM. The penalty parameter and kernel function parameters of SVM are optimized using the PSO. Similarly, Siqiao Yang et al.³² considered a GA for tuning the parameters of SVM for the classification of nine different hand activities.

Recent developments in biomedical signal processing have shown the effectiveness of hybrid techniques combining multiple analytical approaches. For biomedical signal processing, Eraslan et al.³³ successfully used hybrid adaptive neuro-fuzzy inference systems for dynamic time warping-based authentication, while Ozturk et al.³⁴ investigated unified frameworks integrating t-SNE and dynamic time warping for biomedical signal-based biometric applications, offering insightful information for optimizing signal processing frameworks. Also, in parallel, recent advancements in deep learning (DL) approaches have gained significant attention for their ability to extract complex features from biomedical signals and classify gestures. Recent studies^{35,36,37,38,39} explore human-machine interaction and intuitive control in various applications such as gesture recognition, breathing pattern recognition, and gait analysis using DL-based techniques. Moreover, compared to traditional machine learning methods, these recent deep learning models often require large datasets and high-performance computational units which making the systems costlier and bulkier, which may hinder their use in daily-life activities. These recent advancements in biomedical signal processing and machine learning optimization highlight the need for hyperparameter tuning techniques in gesture recognition systems, which aligns with the objectives of this study. In the applications of the biomedical field, a slight change in accuracy is of the utmost value. The precise recognition of hand gestures is also essential because it has numerous applications, such as robotics and prosthesis control. As per the literature survey, there is relatively less work on the tuning of hyperparameters of machine learning models employing an optimization method for hand gesture recognition. Along with the recent growth of research in sEMG-based HGR, relatively few studies have explored the comparative analysis of the advanced metaheuristic techniques for hyperparameter optimization. This study provides a Linear Population Size Reduction Success-History Adaptation Differential Evolution (L-SHADE)-based optimized Extra Tree (ET) MLC framework for HGR. This study also provides a systematic review of ten machine learning models used to recognize six different hand gestures, and then ten distinct optimization techniques are applied to tune the hyperparameters of the machine learning algorithm, which yields the best results.

The main contributions that have been presented in this research article are as follows:

1. 1.

   sEMG signal-based recognition of hand gestures using an optimized extra tree framework by tuning the hyperparameters using L-SHADE optimization.

2. 2.

   L-SHADE optimized extra tree framework is compared with ten optimization approaches for the hyperparameter tuning.

3. 3.

   Comparative analysis of ten machine learning models are applied for the classification of hand gestures.

4. 4.

   sEMG signal of six different hand gestures is acquired from the two forearm muscles of four participants (two males and two females) using BIOPAC MP150.

5. 5.

   Performance parameters of machine learning models without hyperparameter tuning are also examined and compared to the proposed tuned machine learning models.

The following sections comprise this study; a brief introduction is presented in section. In section, the procedure of dataset preparation and feature extraction is presented. Section is divided into ten subsections discussing the machine learning models. Section is divided into ten subsections discussing the optimization technique that is used during the fine-tuning of hyperparameters of machine learning models. Results have been discussed in section . The study is concluded in section with the future scope.

The proposed methodology for classifying hand gesture movements is shown in Fig. 1. It demonstrates the classification stages, including dataset preparation, hyperparameter optimization, and classification using machine learning models. Initially, the forearm muscles’ sEMG data are collected and preprocessed. Relevant features are extracted following data preprocessing and then applied to machine learning models to classify the six hand gesture activities. A high degree of accuracy is most important in the field of medical for mapping assistive devices to tasks similar to natural ones. The performance of machine learning models is highly dependent on their hyperparameters because they control the learning process. To enhance the classification performance, ten distinct optimization algorithms are considered to optimize the parameters of machine learning algorithms. The subsections of the proposed methodology are explained as follows:

Dataset preparation

In this study, data for hand gesture identification are acquired from a group of four willing participants. Prior to their involvement, each participant provided written informed consent, demonstrating their understanding and willingness to take part in the research. The anthropometric data of the participants who participated in the data acquisition process are outlined in Table 1.

The sEMG signals are collected using the AcqKnowledge software version 4.4, which is integrated with the BIOPAC MP150 system, which is well-known for its expertise in EMG data gathering and signal processing. Figure 2 depicts the configuration utilized throughout the data acquisition process. It shows the hardware instrumentation for data acquisition by placing disposable Ag/Ag-Cl surface electrodes on the participant’s forearm muscle to measure the sEMG signal. The extensor digitorum and flexor pollicis longus muscles are specifically targeted for signal assessment^40,41. Based on anatomical landmarks, surface electrodes were placed on the forearm at one-third of the distance between the elbow and the wrist to ensure standardized placement across all subjects^42,43,44. The sEMG signals are captured with a shielded cable electrode coupled to a TEL-100MC filter amplifier set to 1000 gain. This setup provided signal amplification and noise reduction. The data are collected at a sampling rate of 2 KHz to provide high-frequency resolution and a precise representation of muscle activation during hand gesture activities.

Four healthy adult male and female participants signed up for this study. The participants are asked to perform activities in a pre-listed sequence. A total of six activities are considered. Each of these activities is carried out in three stages, and during the task phase, the participant is instructed to meditate and relax.

Step 1: Pre-Task Stage: At this stage, the participants are briefed about the activity two minutes prior to performing the actual task.

Step 2: Task Stage: At this stage, the participants are asked to continuously perform the task for an interval of 40 seconds.

Step 3: Post-Task Stage: After the completion of each activity, the participants are given time to rest. Muscle signals can deteriorate due to fatigue in the muscles.

Preprocessing

The collection of sEMG signals from the muscle is caused due to the action of muscle fibers at the time of muscle contraction. It is affected by unwanted external influences in the form of noise, such as signal line electromagnetic noise, electrode noise, and broadband noise⁴⁵. Therefore, it is quite essential to reduce the noise of the original signal to provide accurate and effective data for feature extraction. Low frequencies of 1-10 Hz, which do not contain important information and are contaminated by movement artifacts, should often be rejected. Power source radiation, often known as Power-Line Interference, is a 50 Hz ambient noise. The influence of this noise can be reduced by using a narrow-band notch filter. Overall, these steps, which serve to reduce noise and prepare the EMG signal estimate, are used in this study:

1. 1.

   Passing the 50 Hz notch filter on the narrowband signal. The notch filter’s transfer function is:

   $$begin{aligned} W(s) = Mfrac{s^4+s^2b_{2}+b_0}{s^4+s^3a_3+s^2a_2+sa_1+a_0} end{aligned}$$

   (1)

   Where (b_n) and (a_n) are the transfer function coefficients, M is the scale transmission coefficient of the filter⁴⁶.

2. 2.

   A (4^{th}) order Butterworth band-pass filter having a cutoff frequency between 10 and 500 Hz. The band-pass filter’s transfer function is:

   $$begin{aligned} W(s) = Hfrac{s^4}{s^4+a_3s^3+a_2s^2+a_1s+a_0} end{aligned}$$

   (2)

   H is an arbitrary multiplicative constant, (a_n) are the transfer function coefficients⁴⁷.

Figure 3 is present as an example of the sEMG signal, so that the impacts of each preprocessing step may be clearly observed. While the figure shows that the raw sEMG signals are noisy, the extreme frequency components are removed following a two-stage filtering process. The preprocessed signal (i.e., Fig. 3) will be used to extract features for machine learning models in the following section.

Feature extraction

The practical units of data from raw signals are known as features. Selecting an appropriate feature is one of the most important factors of classification accuracy. The process of turning raw data into input that the classification system can use is known as feature extraction. The extraction of features from the signal has segmented the signal into intervals of 256 msec windows with an overlapping length equivalent to 25% of the signal⁴⁸. The windowing approach is utilized in this case because the entire data set is too large to analyze and includes redundancy, and the instantaneous EMG signal sample has insufficient information on overall muscle activity. Figure 4 represents windowing with the overlapping segmentation of signals.

To reduce the possibility of reporting unsatisfactory performance in the dataset, extract the most beneficial features for sEMG data based on various characteristics and outcomes. These signal features can be obtained using time domain analysis, frequency domain analysis, or both. In this study, 17 such features have been extracted from each signal segment. Among the features, eleven are extracted in the time domain and six in the frequency domain. The mathematical formulation of features is shown in Table 2.

This study uses ten machine learning models to classify hand gestures. For the employed models, performance parameters are evaluated, and the best classifier is selected to help in the improvement of gesture recognition. For the recognition of hand gestures, machine learning models are as follows: Decision Tree(DT), Random Forest (RF), Adaboost (ADB), Bagging (BAG), Gradient Boosting (GB), Support Vector Machine (SVM), Logistic Regression (LR), Naive Bayes (NB), k-Nearest Neighbour (KNN)^49,50,51, and Extra Tree (ET). The brief description of the most accurate machine learning model, ET is as follows:

Extra Tree (ET)

Extra Tree refers to extremely randomized decision trees. This ensemble has a similar number of decision trees as the random forest but differs in how randomization is incorporated during training. Training and splitting of branches are the primary aspects that set this algorithm apart from the random forest. The ET algortihm uses the classical top-down approach to construct an ensemble of the unpruned decision or regression trees. Its two primary distinctions from previous tree-based ensemble approaches are that it splits nodes by selecting cut-points randomly and grows the trees using the entire learning sample (rather than a bootstrap replica).

The method mainly comprises of two parameters: no. of attributes K randomly selected at each node and the minimal sample size (n_{min}) for splitting a node. Here, attribute refers to a specific input variable used in an ET. The number of trees of the ensemble parameters M, K, and (n_{min}) have different effects: K determines the strength of the attribute selection process, (n_{min}) the strength of averaging output noise, and M the strength of the variance reduction of the ensemble model aggregation. These parameters might be automatically or manually altered to the particulars of the problem. The information gain is specifically normalized in the measured score for classification. This measure can be obtained for a sample S and a split s:

Where (H_c(s)) is the log entropy for the classification, (H_s(S)) is the split entropy, and (I_c^s(S)) is the mutual information of the split outcome and the classification. The algorithm’s remarkable efficiency directly results from these properties^52,53.

Optimization techniques are generally applied for hyperparameter tuning of machine learning models and thus help in improving their performance. These are helpful in finding the optimal parameters for machine learning models. In this study, ten different optimization algorithms have been employed to improve the performance of the extra tree (ET) machine learning model. Moreover, these algorithms have been compared, and the best algorithm among them has been evaluated. Ten different optimization algorithms that have been employed to improve the accuracy of ET are Particle Swarm Optimization (PSO), Atom Search Optimization (ASO), Grey Wolf Optimization (GWO), Genetic Algorithms (GA), Artificial Rabbits Optimization (ARO), Biogeography-Based Optimization (BBO), Artificial Ecosystem-based Optimization (AEO), Harmony Search (HS), Water Cycle Algorithm (WCA)^54,55,56, and Linear Population Size Reduction Success-History Adaptation Differential Evolution (L-SHADE). A brief introduction to the highly dominating L-SHADE algorithm has been mentioned in the following subsections:

Linear population size reduction success-history adaptation differential evolution (L-SHADE)

L-SHADE is one of the state-of-the-art DE algorithms, an adaptive DE (differential evolution) that incorporates the linear success history-based parameter adaption. DE is a more straightforward method that uses a lesser number of parameters. It has been enhanced using the Linear Population Size Reduction is known as LSHADE⁵⁷.

In this algorithm, population is represented as (x_i = (x_1,ldots , x_N)), (i = 1,ldots , n,), where, N is the dimension of the target and n is the population size. It uses Crossover rate CR, scaling factor SF, and a memory set of historical memory cells H containing values (M_{SF, K}, M_{CR, K}) to generate new parameters for the crossover and mutation in every iteration⁵⁸. These parameters are sampled using randomly chosen memory index (k ;epsilon ; [1, H]) as follows:

where randc(a, b) is a random value generated by the Cauchy distribution, and randn(a, b) is a randomly generated value by the normal distribution with position and scale parameters (a, b).

The crossover rate and scaling factor value are in the range [0, 1]. The crossover and scaling factor improvement is stored in the array (S_{SF}), (S_{CR}), together with the fitness difference stored in (Delta f). The updation in the memory cell with index h increasing from 1 to H with every generation is as follows:

where, (w_j=frac{Delta f_{i}}{sum _{k=1}^{left| S right| }Delta f_{k}}) and (Delta f_{j}=left| f(u_j) -f(x_j)right|) and S from the (S_{SF}) or (S_{CR}). The previous parameter values are used to set the new ones with updated parameters c as follows:

where g is the current generation number. In general, the updated parameter c is set at 0.5.

In the L-SHADE algorithm, the number of individuals decreases linearly as it uses linear population size reduction. Therefore, at the end of each iteration, the size of NP is recalculated, and the worst individuals are eliminated from the population. The population size is updated as follows:

where, (NP_{min}), (NP_{max}) are the minimum and maximum population sizes. Similarly, NF, (NF_{max}) are the current and maximum number of function evaluations, respectively.

This section explains the experimental results and their analysis of acquired hand gesture data classification based on the sEMG signal. On the acquired dataset, ten machine learning algorithms have been applied. Initially, the data is split into two parts without disturbing the original sequence. The first part contained 70% of the data from the beginning, which is used during training, and the remaining 30% of the data is reserved for testing. The original sequence is not disturbed to keep the temporal order intact. Four cases of different participants (M1, M2, F1, F2) are considered to generalize the study on classifiers. The four performance indices of these classifiers are recorded as shown in Table 3. From the table, it can be concluded that the ET classifier gives the highest accuracy of around 85.22%, 87.37%, 87.37%, and 76.61% in the case of participants M1, M2, F1, and F2, respectively, as compared to other studied classifiers. The precision corresponding to the same classifier is 86.64%, 87.85%, 87.89%, and 76.72% in the case of participants M1, M2, F1, and F2, respectively. The recall value associated with the ET classifier is 85.21%, 87.37%, 87.37%, and 76.61% in the case of participants M1, M2, F1, and F2, respectively. And the F1-Score is 85.27%, 87.15%, 86.99% and 76.44% in case of participant M11, and F2, respectively. Next to ET, BAG has the maximum accuracy in the case of participant M1, while in the case of participants M2, F1, and F2, LR, GB, and RF have the second-best accuracy, respectively.

Additionally, it can be observed that GB appears at least once among the top three classifiers in all the cases. While LR appears as among the best three classifiers in the case of participants M1, M2, and F1 only. The above-discussed machine learning models have learning hyperparameters with predetermined values. These values of hyperparameters for a classifier with higher accuracy are optimized. Optimization technique with 500 iterations for 30 runs to get a highly accurate model produced by the optimization algorithm to obtain the optimal value of the hyperparameter.

Considerable hyperparameters for the optimal value are provided in Table 4 while tuning the machine learning model. Five tuneable hyper-parameters of the ET algorithm are selected for this purpose.

The first hyperparameter of the extra tree classifier is the number of estimators, which can range from 10 to 100. The number of estimators decides the number of decision trees. The second hyperparameter is the classification criterion. It can be either Gini or Entropy.

The criteria govern the splits in a decision tree. The third hyperparameter is the minimum sample split. Its range is set from 2 to 30; it decides the minimum number of splits of a node. The fourth hyperparameter is the maximum number of features for the selection, which ranges from 2 to 34. The fifth hyperparameter is the model’s maximum depth, which ranges from 1 to 25. It decides the level of tree growth. The hyperparameters are tuned separately for each participant, and the optimal ones are listed in Table 4. Most notably from the table, the Gini criterion appears consistently as the optimal classification criterion parameter in ET. At the same time, the other parameters vary across participants.

Figure 5 shows four boxplots corresponding to four different cases of participants. The boxplot is used to indicate variation in the accuracy of the ET classifier caused by the ten optimizers over the 30 runs. It is observed that classification accuracy remains nearly consistent when L-SHADE is used as an optimizer to tune the hyperparameters of the ET classifier. Also, the consistent value is near to the maximum value of accuracy, which justifies the applicability of the L-SHADE optimizer, as compared to the other optimizers, in this study.

Table 5 represents the improved accuracy obtained after optimally tuning the parameters. Accuracy improvement as seen in the case of ET with L-SHADE is maximum, which is about 88.97%, 90.59%, 91.12%, and 80.91% in the case of participants M1, M2, F1, and F2. The average improvement is about 4% which can provide great relief to users of the prosthetic arm while doing daily chores. Another such application also reduces the chances of mishaps caused during teleoperation. Both of which remain the end effect of this study.

After hyperparameter tuning, the model’s performance measure with and without the optimization is recorded in Table 6. And it can be observed that performance measure accuracy, precision, recall, and F1 score improve with optimization, as well as time is reduced for the participants M1, M2, F1, and F2 when the hyperparameter is optimized.

A confusion matrix illustrates the performance of a classification technique. It contains information about the actual and expected labels that a model evaluates. In the confusion matrix, the diagonal element represents the data point that has been correctly categorized. Figure 6 presents the confusion matrix of the Extra Tree classifier, in which Fig. 6(a)-(d) represents the confusion matrix corresponding to the participants before hyperparameter tuning, while the rest of the subfigures (Fig. 6(e)-(h)) represent the confusion matrix corresponding to the participants after hyperparameter tuning. Figure 6(a) can be interpreted as 44 samples of TE, 59 samples of ME, 53 samples of FME, 47 samples of FMTE, 47 samples of FMRE, and 62 samples of HC are correctly classified, whereas 18 samples of TE, 3 samples of ME, 9 samples of FME, 15 samples of FMTE, 11 samples of FMRE, and zero samples of HC are incorrectly classified. Similarly, observations can be drawn from the rest of the confusion matrices. However, it can be visually compared that the number of correct classifications improves for at least two activities in all cases after tuning.

Assessment of model generalizability

The proposed framework’s generalizability is assessed with the help of validation on the publicly available, more diverse gesture data, and also through the statistical analysis of the model performance through statistical hypothesis testing methods, along with checking the practical viability based on the cross-session performance. The following subsection illustrates the evaluation with a public dataset and the statistical analysis as well on the cross-session viability of the proposed framework.

Evaluation with publicly dataset

The generalizability of the proposed framework, optimized ET with the L-SHADE, was evaluated using the public data repository from R. N. Khushaba et al.⁵⁹, which includes datasets of 15 different classes of hand gestures. During robustness analysis, the same steps are followed, and a similar environment is set up. After ensuring uniformity in the environment, the datasets are analyzed using the two best machine learning models: our proposed model, ET+L-SHADE, and the second most accurate model on the collected data, ET.

From these analyses, it was observed that the proposed architecture achieved the highest accuracy among all subjects (a total of eight). The accuracy attained by the subjects is shown in Fig. 7. The figure clearly shows that the accuracy achieved by subjects s1-s8 with the proposed ET+L-SHADE model had a mean increase in accuracy of approximately 3.0% for all subjects. Thus, in this setup, ET+L-SHADE attained the highest accuracy for both the collected and the publicly available datasets.

Evaluation based on statistical analysis

To validate the significance of the proposed framework, L-SHADE-based optimized Extra Tree MLC results achieved high accuracy. For the statistical significance check, a t-test was conducted on both datasets, acquired. The result of the t-test with the statistic parameter t-value and p-value is illustrated in Table 7. The results on both datasets signify that the p-value is less than 0.05. It means the proposed approach has a significant improvement over the base classifier, and the proposed approach consistently outperforms among both datasets.

Cross-session evaluation

To assess the generalizability of the proposed framework, cross-session variability was checked on the publicly accessible dataset. The dataset comprises recordings from three trial sessions (S1, S2, S3) throughout the acquisition. For the generalizability check, this study assessed that one session was designated for training, while the remaining two sessions were used for testing. This process was repeated for each subject to establish. Such an arrangement replicates a real-world scenario where factors like posture, muscle fatigue, and the placement of the electrode cause disturbances in the signals over time.

The significant results of the cross-session are illustrated in Table 8. The table summarizes the eight subjects’ data performance with respect to the proposed (ET+L-SHADE) in comparison to the ET across various session combinations for individual subjects. The results show outperformance among all the subjects with every session; ET+L-SHADE beats the base classifier ET. Notably, subjects s1, s2, and s5 show gains that are more noticeable in every cross-session compared to the others. These findings highlight the enhanced adaptability and generalizability of the proposed model.

Following the discussion on model generalizability, we elaborate on the computational efficiency to further support the practical applicability of the proposed method. The computational complexity of the L-SHADE-optimized Extra Trees framework is approximately ({mathcal {O}}(N.P.T.n.log n.d)), where N represents the no. of generations, P denotes the size of the population, T is the no. of trees, n signifies the number of training samples, and d is the depth of the tree.

However, once the model is trained, the inference complexity is reduced to ({mathcal {O}}(T.d)) per data sample. This low inference cost makes the proposed framework highly suitable for real-time applications. Furthermore, the Extra Trees classifier supports easy pruning and delivers low-latency decision-making, reinforcing the framework’s practical applicability in real-time prosthetic technologies.

Comparison with state-of-the-art studies

This study introduces an L-SHADE-optimized learning framework for sEMG-based hand gesture recognition (HGR). The superior performance of the proposed framework has been validated using both a self-acquired dataset and a publicly available dataset. Existing literature primarily focuses on improving gesture recognition accuracy, often without optimizing the hyperparameters of machine learning classifiers. This critical aspect is frequently overlooked by researchers. Due to the fact that they are trying to develop a new classifier in place of tuning the parameters of the classifier. To address this gap, we propose an L-SHADE-optimized Extra Tree classifier for enhanced hand gesture recognition. Along with a comparative analysis, a presentation of the ten optimization algorithms that enhance the performance of the subject-specific classifier is provided.

Table 9 summarizes state-of-the-art studies that utilize hyperparameter tuning of machine learning classifiers (MLCs), highlighting their best performances for comparison with our study in terms of key parameters such as no. of channels, no. of classifiers, no. of used optimization techniques, and data for validation. According to the literature, while model performance has generally improved, the aspect of systematic hyperparameter tuning using diverse optimization techniques remains largely underexplored. Therefore, in this study, a detailed comparison is provided with two such studies on the low-end processors and fewer storage units for hardware implementation of a machine learning classifier, which makes the device economical and lighter. One more important thing these machine learning classifiers are easy to interpret and make trustworthy.

1. Venugopal, G., Navaneethakrishna, M. & Ramakrishnan, S. Extraction and analysis of multiple time window features associated with muscle fatigue conditions using semg signals. Expert Systems with Applications 41, 2652–2659 (2014).

   Article  Google Scholar 

2. Zhou, H., Mohammadi, A., Oetomo, D. & Alici, G. A novel monolithic soft robotic thumb for an anthropomorphic prosthetic hand. IEEE Robotics and Automation Letters 4, 602–609 (2019).

   Article  Google Scholar 

3. https://www.who.int/es/news-room/fact-sheets/detail/disability-and-health (2023).

4. Lathan, C. E., Bogner, M. S., Hamilton, D. & Blanarovich, A. Human-centered design of home care technologies. NeuroRehabilitation 12, 3–10 (1999).

   Google Scholar 

5. Alkan, A. & Günay, M. Identification of emg signals using discriminant analysis and svm classifier. Expert systems with Applications 39, 44–47 (2012).

   Article  Google Scholar 

6. Aly, H. & Youssef, S. M. Bio-signal based motion control system using deep learning models: A deep learning approach for motion classification using eeg and emg signal fusion. Journal of Ambient Intelligence and Humanized Computing 14, 991–1002 (2023).

   Article  Google Scholar 

7. Vijayvargiya, A., Prakash, C., Kumar, R., Bansal, S. & Tavares, J. M. R. Human knee abnormality detection from imbalanced semg data. Biomedical Signal Processing and Control 66, 102406 (2021).

   Article  Google Scholar 

8. Wu, H., Li, H., Fang, X. & Luo, X. A survey on teaching workplace skills to construction robots. Expert Systems with Applications 205, 117658 (2022).

   Article  Google Scholar 

9. Reaz, M. B. I., Hussain, M. S. & Mohd-Yasin, F. Techniques of emg signal analysis: detection, processing, classification and applications. Biological procedures online 8, 11–35 (2006).

   Article  Google Scholar 

10. Hakonen, M., Piitulainen, H. & Visala, A. Current state of digital signal processing in myoelectric interfaces and related applications. Biomedical Signal Processing and Control 18, 334–359 (2015).

    Article  Google Scholar 

11. Li, G., Zhang, L., Sun, Y. & Kong, J. Towards the semg hand: internet of things sensors and haptic feedback application. Multimedia Tools and Applications 78, 29765–29782 (2019).

    Article  Google Scholar 

12. Pradhan, A. et al. Acquisition and classification of emg using a dual-channel emg biopotential amplifier for controlling assistive devices. In 2016 IEEE Annual India Conference (INDICON), 1–5 (IEEE, 2016).

13. Pancholi, S. & Joshi, A. M. Portable emg data acquisition module for upper limb prosthesis application. IEEE Sensors Journal 18, 3436–3443 (2018).

    Article  ADS  Google Scholar 

14. Vijayvargiya, A., Singh, P., Kumar, R. & Dey, N. Hardware implementation for lower limb surface emg measurement and analysis using explainable ai for activity recognition. IEEE Transactions on Instrumentation and Measurement 71, 1–9 (2022).

    Article  Google Scholar 

15. Greene, S., Thapliyal, H. & Caban-Holt, A. A survey of affective computing for stress detection: Evaluating technologies in stress detection for better health. IEEE Consumer Electronics Magazine 5, 44–56 (2016).

    Article  Google Scholar 

16. de Souza, J. O. d. O. et al. Investigation of different approaches to real-time control of prosthetic hands with electromyography signals. IEEE Sensors Journal 21, 20674–20684 (2021).

17. Ulkir, O., Gokmen, G. & Kaplanoglu, E. Emg signal classification using fuzzy logic. Balkan Journal of Electrical and Computer Engineering 5, 97–101 (2017).

    Article  Google Scholar 

18. Arteaga, M. V., Castiblanco, J. C., Mondragon, I. F., Colorado, J. D. & Alvarado-Rojas, C. Emg-driven hand model based on the classification of individual finger movements. Biomedical Signal Processing and Control 58, 101834 (2020).

    Article  Google Scholar 

19. Kuzborskij, I., Gijsberts, A. & Caputo, B. On the challenge of classifying 52 hand movements from surface electromyography. In 2012 annual international conference of the IEEE engineering in medicine and biology society, 4931–4937 (IEEE, 2012).

20. Al Omari, F., Hui, J., Mei, C. & Liu, G. Pattern recognition of eight hand motions using feature extraction of forearm emg signal. Proceedings of the National Academy of Sciences, India Section A: Physical Sciences 84, 473–480 (2014).

    Article  ADS  Google Scholar 

21. Song, W. et al. Design of a flexible wearable smart semg recorder integrated gradient boosting decision tree based hand gesture recognition. IEEE transactions on biomedical circuits and systems 13, 1563–1574 (2019).

    Article  PubMed  Google Scholar 

22. Schratz, P., Muenchow, J., Iturritxa, E., Richter, J. & Brenning, A. Hyperparameter tuning and performance assessment of statistical and machine-learning algorithms using spatial data. Ecological Modelling 406, 109–120 (2019).

    Article  Google Scholar 

23. Yang, L. & Shami, A. On hyperparameter optimization of machine learning algorithms: Theory and practice. Neurocomputing 415, 295–316 (2020).

    Article  Google Scholar 

24. Nematzadeh, S., Kiani, F., Torkamanian-Afshar, M. & Aydin, N. Tuning hyperparameters of machine learning algorithms and deep neural networks using metaheuristics: A bioinformatics study on biomedical and biological cases. Computational biology and chemistry 97, 107619 (2022).

    Article  CAS  PubMed  Google Scholar 

25. Chitra, B. & Kumar, S. An optimized deep learning model using mutation-based atom search optimization algorithm for cervical cancer detection. Soft Computing 25, 15363–15376 (2021).

    Article  Google Scholar 

26. Geitle, M. & Olsson, R. A new baseline for automated hyper-parameter optimization. In Machine Learning, Optimization, and Data Science: 5th International Conference, LOD 2019, Siena, Italy, September 10–13, 2019, Proceedings 5, 521–530 (Springer, 2019).

27. Jayaprakash, K. & Balamurugan, S. Artificial rabbit optimization with improved deep learning model for plant disease classification. In 2023 5th International Conference on Smart Systems and Inventive Technology (ICSSIT), 1109–1114 (IEEE, 2023).

28. Lee, W.-Y., Park, S.-M. & Sim, K.-B. Optimal hyperparameter tuning of convolutional neural networks based on the parameter-setting-free harmony search algorithm. Optik 172, 359–367 (2018).

    Article  Google Scholar 

29. Islam, N. & Irshad, K. Artificial ecosystem optimization with deep learning enabled water quality prediction and classification model. Chemosphere 309, 136615 (2022).

    Article  CAS  PubMed  Google Scholar 

30. Raji, I. D. et al. Simple deterministic selection-based genetic algorithm for hyperparameter tuning of machine learning models. Applied Sciences 12, 1186 (2022).

    Article  CAS  Google Scholar 

31. Sui, X., Wan, K. & Zhang, Y. Pattern recognition of semg based on wavelet packet transform and improved svm. Optik 176, 228–235 (2019).

    Article  Google Scholar 

32. Yang, S., Chai, Y., Ai, J., Sun, S. & Liu, C. Hand motion recognition based on ga optimized svm using semg signals. In 2018 11th International Symposium on Computational Intelligence and Design (ISCID), vol. 2, 146–149 (IEEE, 2018).

33. Eraslan, B., Gorur, K. & Temurtas, F. Liveness-verified dynamic time warping-based authentication and hybrid adaptive neuro-fuzzy inference system identification for single-channel diaphragmatic breathing surface electromyography biometrics. Advanced Intelligent Systems 2500015 (2025).

34. Ozturk, H., Eraslan, B. & Gorur, K. Investigation of t-sne and dynamic time warping within a unified framework for resting-state and minor analysis visual task-related eeg alpha frequency in biometric authentication: A detailed analysis. Digital Signal Processing 160, 105042 (2025).

    Article  Google Scholar 

35. Xiong, D., Zhang, D., Chu, Y., Zhao, Y. & Zhao, X. Intuitive human-robot-environment interaction with emg signals: a review. IEEE/CAA Journal of Automatica Sinica 11, 1075–1091 (2024).

    Article  Google Scholar 

36. Xiong, D., Zhang, D., Zhao, X. & Zhao, Y. Deep learning for emg-based human-machine interaction: A review. IEEE/CAA Journal of Automatica Sinica 8, 512–533 (2021).

    Article  Google Scholar 

37. Xiong, D., Zhang, D., Zhao, X. & Zhao, Y. Hand gesture recognition using instant high-density emg graph via deep learning method. In 2020 Chinese Automation Congress (CAC), 5143–5148 (IEEE, 2020).

38. Xiong, D., Zhang, D., Zhao, X., Chu, Y. & Zhao, Y. Synergy-based neural interface for human gait tracking with deep learning. IEEE Transactions on Neural Systems and Rehabilitation Engineering 29, 2271–2280 (2021).

    Article  PubMed  Google Scholar 

39. Xiong, D., Zhang, D., Zhao, X., Chu, Y. & Zhao, Y. Breathing pattern recognition by the fusion of emg and acceleration signals. In 2022 IEEE International Conference on Robotics and Biomimetics (ROBIO), 877–882 (IEEE, 2022).

40. Shi, W.-T., Lyu, Z.-J., Tang, S.-T., Chia, T.-L. & Yang, C.-Y. A bionic hand controlled by hand gesture recognition based on surface emg signals: A preliminary study. Biocybernetics and Biomedical Engineering 38, 126–135 (2018).

    Article  Google Scholar 

41. Hamrick, M. W., Churchill, S. E., Schmitt, D. & Hylander, W. L. Emg of the human flexor pollicis longus muscle: implications for the evolution of hominid tool use. Journal of Human Evolution 34, 123–136 (1998).

    Article  CAS  PubMed  Google Scholar 

42. Guo, W. et al. Hand kinematics, high-density semg comprising forearm and far-field potentials for motion intent recognition. Scientific Data 12, 445 (2025).

    Article  PubMed  PubMed Central  Google Scholar 

43. Mongardi, A. et al. Hand gestures recognition for human-machine interfaces: A low-power bio-inspired armband. IEEE Transactions on Biomedical Circuits and Systems 16, 1348–1365 (2022).

    Article  PubMed  Google Scholar 

44. Fan, X., Rydgård, J., Yang, L. & Johansson, P. J. Assessing forearm exertion in manual tasks with surface emg: A comparative analysis of through-forearm vs. muscle specific emg placements. bioRxiv 2024–12 (2024).

45. Chowdhury, R. H. et al. Surface electromyography signal processing and classification techniques. Sensors 13, 12431–12466 (2013).

    Article  ADS  PubMed  PubMed Central  Google Scholar 

46. Geiger, R. L. & Sanchez-Sinencio, E. Active filter design using operational transconductance amplifiers: A tutorial. IEEE circuits and devices magazine 1, 20–32 (1985).

    Article  Google Scholar 

47. Su, K. L. Analog filters (Springer Science & Business Media, USA, 2012).

    Google Scholar 

48. Vijayvargiya, A. et al. Hybrid deep learning approaches for semg signal-based lower limb activity recognition. Mathematical Problems in Engineering 2022 (2022).

49. Gehlot, N., Jena, A., Vijayvargiya, A. & Kumar, R. Surface electromyography based explainable artificial intelligence fusion framework for feature selection of hand gesture recognition. Engineering Applications of Artificial Intelligence 137, 109119 (2024).

    Article  Google Scholar 

50. Gehlot, N., Jena, A., Vijayvargiya, A. & Kumar, R. semg-based classification of finger movement with machine learning. In 2023 International Conference on Computer, Electronics & Electrical Engineering & their Applications (IC2E3), 1–6 (IEEE, 2023).

51. Saxena, S., Yadav, S., Singh, B., Kumar, R. & Chaudhary, S. Machine learning-based recognition of white blood cells in juvenile visayan warty pigs. In 2023 International Conference on Artificial Intelligence for Innovations in Healthcare Industries (ICAIIHI), vol. 1, 1–6 (IEEE, 2023).

52. Geurts, P., Ernst, D. & Wehenkel, L. Extremely randomized trees. Machine learning 63, 3–42 (2006).

    Google Scholar 

53. Vijayvargiya, A., Kumar, R., Dey, N. & Tavares, J. M. R. Comparative analysis of machine learning techniques for the classification of knee abnormality. In 2020 IEEE 5th International Conference on Computing Communication and Automation (ICCCA), 1–6 (IEEE, 2020).

54. Jacob, I., Lamba, R., Kumar, R. & Montero, F. J. Metaheuristic based single and multiobjective optimization of thermoelectric generator. Applied Thermal Engineering 236, 121790 (2024).

    Article  CAS  Google Scholar 

55. Van Thieu, N. & Mirjalili, S. Mealpy: An open-source library for latest meta-heuristic algorithms in python. Journal of Systems Architecture 139, 102871 (2023).

    Article  Google Scholar 

56. Mehta, A., Jangid, J., Saxena, A., Shekhawat, S. & Kumar, R. Harmonic estimator using design atom search optimization algorithm. In Proceedings of International Conference on Communication and Computational Technologies: ICCCT 2021, 181–198 (Springer, 2021).

57. Tanabe, R. & Fukunaga, A. S. Improving the search performance of shade using linear population size reduction. In 2014 IEEE congress on evolutionary computation (CEC), 1658–1665 (IEEE, 2014).

58. Al-Dabbagh, R. D., Neri, F., Idris, N. & Baba, M. S. Algorithmic design issues in adaptive differential evolution schemes: Review and taxonomy. Swarm and Evolutionary Computation 43, 284–311 (2018).

    Article  Google Scholar 

59. Khushaba, R. N. & Kodagoda, S. Electromyogram (emg) feature reduction using mutual components analysis for multifunction prosthetic fingers control. In 2012 12th International Conference on Control Automation Robotics & Vision (ICARCV), 1534–1539 (IEEE, 2012).

60. Prakash, K. S. & Kunju, N. An optimized electrode configuration for wrist wearable emg-based hand gesture recognition using machine learning. Expert Systems with Applications 274, 127040 (2025).

    Article  Google Scholar 

61. Rezaee, K., Khavari, S. F., Ansari, M., Zare, F. & Roknabadi, M. H. A. Hand gestures classification of semg signals based on bilstm-metaheuristic optimization and hybrid u-net-mobilenetv2 encoder architecture. Scientific Reports 14, 31257 (2024).

    Article  ADS  CAS  PubMed  PubMed Central  Google Scholar 

62. Asif, A. R. et al. Performance evaluation of convolutional neural network for hand gesture recognition using emg. Sensors 20, 1642 (2020).

    Article  ADS  PubMed  PubMed Central  Google Scholar 

63. Bai, D., Liu, T., Han, X. & Yi, H. Application research on optimization algorithm of semg gesture recognition based on light cnn+ lstm model. Cyborg and bionic systems (2021).

64. Sakinala, U. C. & Abinaya, S. Enhanced detection of hand gestures from semg signals using stacking ensemble with particle swarm optimization and meta-classifier. IEEE Access (2025).

65. Xiong, D. et al. Robotic telemanipulation with emg-driven strategy-assisted shared control method. Science China Technological Sciences 67, 3812–3824 (2024).

    Article  ADS  Google Scholar 

66. Xiong, D., Zhang, D., Zhao, X., Chu, Y. & Zhao, Y. Learning non-euclidean representations with spd manifold for myoelectric pattern recognition. IEEE Transactions on Neural Systems and Rehabilitation Engineering 30, 1514–1524 (2022).

    Article  PubMed  Google Scholar 

Download references

Author notes

1. Naveen Gehlot

   Present address: Department of Instrumentation and Control Engineering, Manipal Institute of Technology, Manipal Academy of Higher Education, Manipal, India

Authors and Affiliations

1. Department of Electrical Engineering, Malaviya National Institute of Technology, 302017, Jaipur, Rajasthan, India

   Naveen Gehlot, Ashutosh Jena, Rajesh Kumar & Surender Hans

2. Dhirubhai Ambani Institute of Information and Communication Technology, Gandhinagar, India

   Ankit Vijayvargiya

3. Department of Human Anatomy and Physiology, Faculty of Health Sciences, University of Johannesburg, 2094, Johannesburg, South Africa

   Rajesh Kumar

4. Department of Mathematics, Malaviya National Institute of Technology, 302017, Jaipur, Rajasthan, India

   Priyanka Harjule

Contributions

All the authors contributed to each part of this paper equally.

Corresponding authors

Correspondence to Naveen Gehlot or Rajesh Kumar.

Ethical approval

This study was performed in accordance with the Nuremberg Code. This human study was approved by MNIT Jaipur. All adult participants provided written informed consent to participate in this study.

Competing interests

The authors declare no competing interests.

Consent to participate

All authors declare that they have the consent to participate.

Consent to publish

All authors declare that they have consent to publish.

Publisher’s note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.

Reprints and permissions