Machine Learning-Based Classification of Body Imbalance and Its Intensity Using Electromyogram and Ground Reaction Force in Immersive Environments

Abstract

Body balancing is a complex task that includes the coordination of muscles, tendons, bones, ears, eyes, and the brain. Imbalance or disequilibrium is the inability to maintain the center of gravity. Perpetuating body balance plays an important role in preventing us from falling or swaying. Biomechanical tests and video analysis can be performed to analyze body imbalance. The musculoskeletal system is one of the fundamental systems by which our balance or equilibrium is sustained and our upright posture is maintained. Electromyogram (EMG) and ground reaction force (GRF) monitoring can be utilized in cases where a rapid response to body imbalance is a necessity. Body balance also depends on visual stimuli that can be either real or virtual. Researchers have used virtual reality (VR) to predict motion sickness and analyze heart rate variability, as well as in rehabilitation. VR can also be used to induce body imbalance in a controlled way. In this research, body imbalance was induced in a controlled way by playing an Oculus game and, simultaneously, EMG and GRF were recorded. Features were extracted from the EMG and were then fed to a machine learning algorithm. Several machine learning algorithms were tested and upon 10-fold cross-validation; a minimum accuracy of 71% and maximum accuracy of 98% were achieved by Gaussian Naïve Bayes and Gradient Boosting classifiers, respectively, in the classification of imbalance and its intensities. This research can be incorporated into various rehabilitative and therapeutic systems.

1. Introduction

Worldwide, falls rank as the second most common cause of fatal accidental injuries. Over 80% of the 684,000 people who die from falls worldwide each year do so in low- and middle-income nations. Most fatal falls occur in adults over the age of 60. Every year, 37.3 million falls happen that are serious enough to need medical care [1,2]. Body balance, also called body equilibrium, is the ability of our body to perpetuate the center of gravity, making our daily life tasks easier to perform and protecting us from injuries. Body equilibrium is maintained through the coordination of complex sensory and motor systems, also called sensorimotor systems [3]. The intricate connections between different sensory inputs, central nervous system processing, and motor responses are part of the sensorimotor systems that keep the body balanced. These systems include the somatosensory system, which incorporates proprioceptors in muscles and joints to provide input on body position; the visual system, which provides spatial orientation; and the vestibular system, which senses head position and movement. While the central nervous system combines sensory data and produces motor orders, the muscular system, especially in the lower limbs and core, reacts to signals from the neurological system [3,4]. Balance is influenced by proprioception and reflexes, and the cerebellum is essential for motor coordination. The proper operation of these elements is guaranteed by the postural control system, which enables people to adapt and stay stable while engaging in a variety of activities. Dysfunction in any of these systems can lead to balance disorders and an increased risk of falls [5].

Electromyography is a method to acquire the electrical activity of muscles, named an electromyogram or EMG. EMG gives crucial insights about the activation patterns of the muscles that contribute to maintaining body balance. Those activation patterns can be monitored using both invasive and non-invasive techniques. Needle electrodes are used in invasive monitoring of EMG, limited to specific motor units, causing pain and discomfort, which may later cause an infection [6]. On the other hand, surface electrodes are easier to apply on the skin and cover a larger area of muscle from the surface. Noise can be effectively removed using filters specific to EMG [7,8]. The use of needle electrodes may cause injury during studies related to human movements [9]. Therefore, using surface electrodes to acquire EMG is the optimal option for conducting studies related to body balancing and movements. By using EMG, one can evaluate muscle activity and neuromuscular control, which is important for determining how the body stabilizes and interacts with imbalance stimuli [8,10,11].

Artificial imbalance can be engineered to replicate real-world conditions, supporting studies on balance problems, sports training, and rehabilitation [12,13,14]. By using virtual spaces to stimulate the sensory and motor systems, people may adjust and enhance their reactions to balance in a supervised and safe environment [15]. This use of VR has the potential to improve postural control evaluation, balance rehabilitation methods, and our knowledge of how the human body adjusts to controlled imbalance situations [16,17,18]. Ground reaction force (GRF) provides information on the strength and direction of forces applied to the ground during actions like walking and standing. Force platforms can be used to estimate the center of pressure (COP) using the values of GRF [19,20]. It can be used along with a head-mounted display to evaluate the changes in COP due to visual stimuli, induced virtually [21,22,23]. It has been observed that using surfacing electromyogram (sEMG) signals along with the data from mechanical sensors improves the overall accuracy of the study compared to using only mechanical sensor data [24]. EMG-powered prosthetics are widely used in the field of rehabilitation. They can be used to control lower limb prosthetics independent of the gait phase [25]. Features extracted from EMG are also being used for classification purposes of locomotion modes and gait phases with the application of machine and deep learning techniques [26,27]. An average accuracy of 87.4% was achieved using artificial neural networks in the classification of stance and swing phases based on EMG signals [28].

In a study, wearing a head-mounted display (HMD) and watching one of the three scenes in a randomized order, participants completed four standing activities in three distinct situations. During their standing duties, they had to keep their arms at their sides and try their best to maintain their stability. A virtual reality-based tracker placed in the lumbar region was used to capture the body wobble. When performing tandem and one-leg standing tasks, the effects of visual sights on postural stability interact considerably. During these tasks, there are larger sway parameters in the VrS and BS conditions compared to the RS condition and lower BS/VrS ratios compared to BS/RS ratios [29]. Similarly, in another study, when the virtual moving room traversed in the participant’s direction and caused an unanticipated movement, it caused noticeably more postural sway than when it moved in the opposite direction and caused an expected movement. The results of the investigation showed a considerable improvement in confidence in balancing without simulator sickness and with genuine immersion. The virtual moving room may create a lifelike virtual reality experience without hurting confidence in balancing or generating VR simulator sickness [30].

To create a VR environment that consistently illustrates the reweighting of sensory information across the vestibular, somatosensory, and visual systems, Virtual Reality Comprehensive Balance Assessment and Training (VR-ComBAT) was used. VR-ComBAT can identify multidirectional shifts in young, healthy participants’ postural control in terms of the center of pressure (COP) under various circumstances. To detect the displacement and velocity of COP, VR-ComBAT is composed of a head-mounted device with integrated VR, two VR tracking sensors, and a force plate connected to the VR headset. For every VR-ComBAT exercise, participants were instructed to tread on the force plate or the foam surface and hold their posture. The study’s objectives were to evaluate VR-ComBAT’s safety, sensation of presence, system use, and face validity in young, healthy subjects [31]. In another study, automated terrain classification in lower limb rehabilitation systems was presented using both the EMG and GRF signals to achieve a high accuracy of 96.8%. The researchers collected EMG and GRF signals from ten healthy subjects and extracted 21 statistical features. A support vector machine (SVM) machine learning model was used for the application of smart neuromuscular-mechanical fusion and machine learning-based terrain classification techniques [32].

Hence, from [27,28,29,30,31,32], a clear research gap can be observed in terms of monitoring GRF and EMG while inducing controlled body imbalance in an immersive environment and then classifying the body imbalance intensity into low, medium, and high. Therefore, using Oculus Quest 2 to induce controlled body imbalance and establish an immersive environment, as well as using a force platform and BIOPAC MP 36 to record resulting variations in the GRF and EMG signals and classify body imbalance and its intensities using machine learning techniques, was considered an optimal option to fill that research gap. Machine learning techniques are used to classify body imbalance into low, medium, and high intensities.

2. Materials and Methods

2.1. Participants

Students at Air University were the participants in this research. A formal consent form was filled out by each participant before data collection. Data were collected from a total of six students. All six of the students were male and aged between 20 and 23 years without any previous history of neuromuscular disorder. They had never undergone balance testing before, ensuring unbiased data collection. The participants were first told about the use of Oculus and its controllers to play the game, electrode placement for recording EMG, and the position of the feet while standing on the force platform to record GRF. Only healthy participants were included in this study by monitoring their physiological parameters like blood pressure, oxygen saturation, and heart rate before starting the data collection. They participants also did not have any hair on their skin to make sure that the EMG signal was free of noise. Participants with a history of psychological problems and general health issues were excluded. As mentioned in

Table 1. Physiological parameters’ range for inclusion criteria of participants for the data collection.

Sr. No.	Physiological Parameter	Normal Range for Inclusion
1.	Blood Pressure	Systolic Pressure	90–120 mm of Hg
		Diastolic Pressure	60–80 mm of Hg
2.	Oxygen Saturation	>95%
3.	Heart Rate	60–100 Beats Per Minute

, only participants who had normal values of physiological parameters like blood pressure, oxygen saturation, and heart rate, without any previous history medical complications, were included.

2.2. Experimental Setup

The BIOPAC MP 36 system was utilized in the investigation to obtain thorough electromyograms (EMGs) from both legs. Surface electrodes were placed on the vastus lateralis muscles of both legs to acquire EMG signals of the left and right legs. Ground reaction force (GRF) measurements were simultaneously acquired with the PASCO force platform. GRFs of both sides of the legs were added up, resulting in only two GRFs, i.e., left leg GRF and right leg GRF. With Oculus Quest 2, immersive settings that induce body imbalance were produced. For this purpose, the Epic Roller Coasters game was selected to induce body imbalance in a controlled way. Figure 1 shows the experimental setup of this research.

2.3. Procedure

A thorough briefing about the paradigm of the study was given to the participants before starting data collection. The participants were also first given training to use Oculus Quest 2. The Epic Roller Coasters game was used to create controlled body imbalance, and while participants were performing the activities that caused the imbalance, real-time EMG and GRF responses were being collected. The virtual roller coaster ride was very effective in inducing imbalance in the participant’s body as well as ensuring his safety as the controllers were in the hands of the participant; hence, the entire game could be terminated whenever he wanted based on his comfort level. Figure 2 shows the complete workflow of this research.

2.4. Data Collection

Throughout the controlled imbalance activities, the targeted lower limb muscles’ EMGs were recorded and compared. Amongst the anterior, posterior, and lateral muscles of the thigh, lateral muscles contribute more towards stabilizing the body during body imbalance. It was observed that the vastus lateralis muscle, the lateral thigh muscle, also had the maximum amplitude of EMG compared to the other muscles of the lower limb, like the gastrocnemius, tibialis anterior, and peroneus muscles. Simultaneously, GRF data were recorded to examine the forces applied to the ground in body imbalance while the participant was playing the virtual game in an immersive environment. The participants played the virtual game for 5 min, or 300 s. It has been observed that the participants gained more control over their body balance in the last one-and-a-half minutes of the gameplay. The BIOPAC MP 36 was set in a frequency range of 30–1000 Hz to acquire the EMG signals from the vastus lateralis muscle of both legs. Also, a band stop or notch filter of 50 Hz was configured in the software to reduce the power line noise. The EMG signals from BIOPAC MP 36 were acquired at a sampling rate of 1000 Hz. GRF signals were acquired using the PASCO force platform at a sampling rate of 1000 Hz, as shown in Figure 3.

2.5. Virtual Reality Setup

An immersive virtual world was established using Oculus Quest 2, which also served as a platform for creating regulated body imbalance using the Rock Falls mode of a game named Epic Roller Coasters, as shown in Figure 4. This arrangement made it possible to precisely control the sensory input, closely mimicking actual situations to observe participants’ responses [33].

2.6. Feature Extraction

Firstly, the acquired EMG and GRF signals of both legs were normalized [34,35,36]. Both time and frequency domain features were extracted. GRF was used for thresholding the imbalance intensities into low imbalance, medium imbalance, and high imbalance. It can be seen in Figure 5 that both the EMG and GRF change whenever the participant is subjected to induced body imbalance.

Epochs of 2 s were formed and both EMG and GRF were plotted in the process of making an envelope. Six features were extracted for each leg, namely mean, standard deviation, skewness, kurtosis, dominant frequency, and force envelope.

2.6.1. Mean

It is important to find the mean value as it gives insights into the central point of the data. The mean value of the EMG was calculated for each epoch having a time duration of 2 s. The mean value of EMG can be calculated by adding up all the values of EMG in the epoch and dividing it by the total number of the values. Mathematically, mean μ can be calculated as

$$\mu ={\sum }_{i=1}^n{\displaystyle \frac{Xi}{N}}$$

(1)

where Xi is the value of EMG and N is the total number of sample values.

2.6.2. Standard Deviation

Standard deviation shows the variation from the central mean point of the data. It tells us about the dispersion of data concerning the mean. Mathematically, standard deviation σ can be calculated as

$$\sigma ={\displaystyle \frac{\sqrt{{\sum }_1^n{(Xi-\mu )}^2}}{N}}$$

(2)

where Xi is the value of EMG, μ is the mean value, and N is the total number of sample values.

2.6.3. Skewness

Skewness is the measure of asymmetry within a dataset. Skewness shows the horizontal spread of the data in a normal distribution. It can be calculated as

$$SK={\displaystyle \frac{{\sum }_1^n{(Xi-\mu )}^3}{\left(N-1\right){\mathsf{\sigma }}^3}}$$

(3)

where Xi is the value of EMG, μ is the mean value, σ is the standard deviation, and N is the total number of sample values.

2.6.4. Kurtosis

Kurtosis shows the normal distribution of data around the mean values. It also refers to the vertical spread of the data in a normal distribution. It can be calculated as

$$Kurt={\displaystyle \frac{{\sum }_1^n{(Xi-\mu )}^4}{N{\mathsf{\sigma }}^4}}$$

(4)

where Xi is the value of EMG, μ is the mean value, σ is the standard deviation, and N is the total number of sample values.

2.6.5. Dominant Frequency

The frequency with maximum energy amongst all the frequencies present in the power spectrum is called the dominant frequency of a signal. We can easily find the dominant frequency by taking the Fourier transform of the time domain signal and then comparing the values of the energies of different frequencies.

2.6.6. Force Envelope

Force envelopes were calculated using the root mean square values of the GRF data. Force envelope (FE) can be calculated as

$$FE={\displaystyle \frac{\sqrt{{\sum }_1^n{\left(GRF\right)}^2}}{N}}$$

(5)

where GRF is the value of ground reaction force and N is the total number of sample values.

2.7. Machine Learning Classification

After preprocessing and feature extraction, EMG signals can be used for classification through the application of different machine learning classifiers [37]. GRF and EMG data were processed to extract features, and machine learning techniques were used to classify imbalance and its intensity, as shown in Figure 6. K-fold cross-validation is an important method to evaluate the performance of machine learning classifiers. It shows how well the classifiers perform on unseen data. In this study, 10-fold cross-validation was used to assess the accuracy of the trained machine learning model.

The opted methodology has been proven to be very effective in the classification of body imbalance and its intensities into low, medium, and high. The results showed this method to have the potential to be integrated into various real-time systems in the domains of rehabilitation, sports medicine, physical therapy, and gaming.

3. Results

3.1. Power Spectral Analysis of EMG

EMG of different muscles of the lower limbs was acquired for testing the muscles that respond with high amplitudes during body imbalance. Four of the muscles were tested, namely vastus lateralis, gastrocnemius, tibialis anterior, and peroneus muscles. It was observed that the vastus lateralis produced EMG signals with high amplitude. The electrode placement, i.e., positive, negative, and ground, is shown in Figure 7A while the subject is standing on the PASCO force platform and electrode leads are connected to the BIOPAC MP 36. A small chunk of the EMG signal is shown in Figure 7B, showing the activity of the vastus lateralis muscle of one of the lower limbs of the subject. Figure 7C–F show the results of different frequency and power spectral analysis techniques, providing insights into the power spectrum and its density of the time domain EMG signal as well as the frequency components of the signal contributing the most during the body imbalance.

3.2. Ten-Fold Cross Validation

The average 10-fold cross-validation accuracies were computed. K-fold cross-validation is one of the methods to validate the results, as it shows how well the classifier performs on unseen data. Hence, its average accuracy holds importance by showing that the trained model performs well on both the seen and unseen data effectively. The average 10-fold cross-validation accuracies are shown in Figure 8A. Upon 10-fold cross-validation, accuracy of 90%, 87%, 71%, 97%, 98%, 98%, 96%, 69%, and 97% was achieved by k-Nearest Neighbor, Linear Discriminant Analysis, Gaussian Naïve Bayes, Decision Tree, Random Forest, Gradient Boosting, Multilayer Perceptron, AdaBoost, and Bagging classifiers, respectively, in the classification of imbalance and its intensities. The error bars were used to plot the standard deviation of the classifiers in computing average 10-fold cross-validation accuracy. It shows the variation in the data and measures the uncertainty or error in the results. Hence, it can be seen in Figure 8B that the error in computation is negligible in the case of Decision Tree, Random Forest, Gradient Boosting, and Bagging classifiers.

3.3. Confusion Matrices

Some of the obtained confusion matrices are shown in Figure 9A–H, where high, medium, and low are the intensities of the body imbalance. The confusion matrix evaluates the performance of a classifier and represents it in the form of a summary by comparing predicted and true labels. In case of body imbalance intensity classification into high, medium, and low, confusion matrices obtained were of multi-class classification. Therefore, all the diagonal values are true positives for the corresponding class. The false positives in the confusion matrices are those that are predicted positive but belong to any of the remaining two classes, whereas the false negatives are those that are predicted negative incorrectly and should have been predicted positive, as shown in Figure 9A–H.

3.4. Performance Metrics

Different evaluation metrics like accuracy, precision, recall, and F1 score were computed using the obtained confusion matrices as shown in Figure 10A, and the corresponding p-value obtained from the t-Test between the test accuracies and cross-validation accuracies for each classifier to check the overfitting amongst the classifiers is shown in Figure 10B. Decision Tree, Random Forest, Gradient Boosting, Multilayer Perceptron, and Bagging classifiers achieved maximum accuracy, precision, recall, and F1 score, whereas the Gaussian Naïve Bayes classifier achieved minimum accuracy, precision, recall, and F1 score amongst all other classifiers. The accuracy shown in this graph is without the k-fold cross-validation. A p value less than 0.05 is the indicator of potential overfitting. The p-value in the case of Random Forest is less than 0.05, and hence there is overfitting in that case, whereas there is no overfitting in the case of other classifiers. All the classifiers except Random Forest performed exceptionally well in the classification of body imbalance and its intensities into low, medium, and high.

4. Discussion

Body imbalance contributes significantly to the variation in the patterns of both the GRF and EMG. Whenever body imbalance is induced, maintaining the center of gravity as well as the activation of muscles responsible for the stabilization of the body with respect to the center of gravity is crucial in avoiding a fall. The vastus lateralis muscle significantly contributes to balancing the body as it was observed that the vastus lateralis muscle had the maximum amplitude of EMG compared to the other muscles of the lower limb, like the gastrocnemius, tibialis anterior, and peroneus muscles, as shown in Figure 7. Oculus Quest 2 was used to induce body imbalance effectively by creating virtual reality immersion using a game called Epic Roller Coasters (Rock Falls Mode), as shown in Figure 4. The variations in GRF and EMG signals occur simultaneously, as shown in Figure 5. Using GRF and EMG to signals simultaneously forms a firm basis to classify imbalance intensities into high, medium, and low. Both normal and 10-fold cross-validation evaluation metrics showed the effectiveness of the machine learning classifiers as shown in Figure 8, Figure 9 and Figure 10.

Upon 10-fold cross validation, it has been observed that the standard deviation shown using the error bar plot in Figure 8B is minimum for Decision Tree, Random Forest, Gradient Boosting, and Bagging classifiers. These classifiers achieved the highest average 10-fold cross validation accuracies. The performance metrics were evaluated using the obtained confusion matrices. The values of true positives, true negatives, false positives, and false negatives from confusion matrices were used to compute accuracy, precision, recall and F1 score for all machine learning classifiers, as shown in Figure 9 and Figure 10.

Higher accuracy does not always indicate optimal performance of any classifier. Some classifiers tend to overfit the data, and when tested on unseen data, their performance drops and they fail perform well in classification. For this purpose, we must check the classifiers for potential overfitting of the data. One of the ways to test the overfitting of the data is to perform a statistical test named the t-Test. This test can be performed between training and test accuracies obtained from the classification of a machine learning classifier. If the p-value after conducting the t-Test is less than 0.05, this indicates that the classifier is overfitting the data. As shown in Figure 10, only the p-value for the Random Forest classifier is less than 0.05, which means the Random Forest classifier is overfitting the data.

This study is only limited to muscles of the lower limb, whereas various muscles of the back can also be used to classify body imbalance and its intensities. Also, this study was conducted on only six healthy and young male subjects. It can also be conducted on both male and female healthy subjects as well as elderly subjects suffering from body imbalance impairments. Different virtual reality immersion environments can be tested to further optimize this work. Other physiological recordings like Electroencephalogram (EEG) and Functional Near Infra-Red Spectroscopy (f-NIRS) can also be incorporated into this experimental setup to compare the efficiency of different modalities in the detection and classification of body imbalance, as well as to enhance the feedback mechanisms in the process of rehabilitation, physical therapy, sports medicine, and patient recovery assessments.

Research of this kind has a lot of potential to be translated into clinical and social applications. From the analysis of the GRF and EMG data, clinicians can proactively treat patients who have impaired balance control. It can be valuable for designing targeted rehabilitation protocols regarding balance and fall prevention in elderly subjects. It is possible to use the classification of the intensities of imbalance to indicate the approaches that can be used in rehabilitation. For instance, patients with high-intensity imbalance can be trained more intensively in terms of strength and stability, whereas patients with medium or low levels of imbalance should be trained less intensively. Its application can be highly beneficial for patients and help physical therapy to be as effective as possible in terms of recovery. The recording of GRF and EMG signals can give clear readings throughout the rehabilitation process of a patient. Health care professionals can monitor the change in balance performance, either in the short or long term, depending on real-time data of muscle activation patterns and ground reaction forces to modify treatment strategies. Likewise, in sports, balance and stability play a major role in a competitor’s performance. GRF and EMG data enable athletes to diagnose motor behaviors, determine their areas of weakness, and plan strategies for improvement. Applications of these kinds of data can go a long way in helping athletes increase their capability and, at the same time, reduce the prevalence of such injuries. Balance analysis using GRF and EMG can help inform ergonomic improvements in the workplace. Employers can increase worker stability and reduce the risk of musculoskeletal injuries by identifying tasks with high imbalance intensities and modifying work settings or implementing training programs. Data from GRF and EMG investigations can help to shape public health campaigns aimed at enhancing balance and mobility in varied populations. For example, community programs aimed at increasing physical activity in older individuals can use the data from these studies to construct effective balance training exercises.

5. Conclusions

Classification of body imbalance and its intensities has the potential to revolutionize adaptive technologies, rehabilitation techniques, exoskeletons, and biomechanical analysis. Visual stimuli using immersive technology like virtual reality can be used to induce controlled body imbalance and, at the same time, electromyogram and ground reaction force signals can be recorded. It is important to note that such hardware and software can be incorporated into real-time rehabilitation systems to enhance their efficiency. Furthermore, it can be used for practical purposes to enhance the stability of athletes going through physical therapy and sports medicine. It can also be used to evaluate the medical procedures and recovery of patients suffering from body imbalance.