Automatic Assessment of Abdominal Exercises for the Treatment of Diastasis Recti Abdominis Using Electromyography and Machine Learning

Abstract

Diastasis Recti Abdominis (DRA) is a medical condition in which the two sides of the rectus abdominis muscle are separated by at least 2.7 cm. This happens when the collagen sheath that exists between the rectus muscles stretches beyond a certain limit. The recti muscles generally separate and move apart in pregnant women due to the development of fetus in the womb. In some cases, this intramuscular gap will not be closed on its own, leading to DRA. The primary treatment procedures of DRA involve different therapeutic exercises to reduce the inter-recti distance. However, it is tedious for the physiotherapists to constantly monitor the patients and ensure that the exercises are being done correctly. The objective of this research is to analyze the correctness of such performed exercises using electromyogram (EMG) signals and machine learning. To the best of our knowledge, this is the first work reporting the objective evaluation of rehabilitation exercises for DRA. Experimental studies indicate that the surface EMG signals were effective in classifying the correctly and incorrectly performed movements. An extensive analysis was carried out with different machine learning models for classification. It was inferred that the RUSBoosted Ensembled classifier was effective in differentiating these movements with an accuracy of 92.3%.

1. Introduction

Diastasis Recti Abdominis (DRA) is the divarication of the rectus abdominis muscle caused by the stretching and widening of the linea alba, resulting in increased inter-recti distance. DRA occurs due to the mechanical and functional disturbances in the anterior abdominal wall. It occurs when the recti muscles separate beyond a limit and fail to come back closer together even eight weeks after pregnancy. Today, DRA exists in 33% of women twelve weeks post-pregnancy. DRA is most commonly seen in postpartum women. It brings about extreme discomfort to those experiencing it. Many women state that their bodies do not feel like they used to prior to pregnancy, and they tend to experience a fear of movement owing to an unstable feeling in the mid-section of their abdomen [1]. Bodily dissatisfaction is a common effect of DRA. Many also experience difficulty while lifting heavy objects while some extreme cases may also develop hernia as a complication. Seven randomized control trials of 381 women found that exercises that specifically targeted only the pelvic floor muscles did not have much effect on bringing about linea alba reduction postpartum [2]. The condition is initially diagnosed through a physical examination carried out by qualified medical professionals. An abdominal ultrasonography, a scan of the abdominal region, is also a common mode of diagnosis of DRA.

Many technological advancements are being made in the field of medicine. Starting from the virtual diagnosis of common ailments to robotic surgeries, automation plays a major role in today’s healthcare as it reduces the workload of medical professionals by decreasing human intervention. With respect to physical and muscular ailments, physiotherapy and muscle activation exercises are most essential to ensuring recovery. It is a cumbersome task for the physiotherapists and doctors to constantly monitor a patient and ensure that they are performing an exercise correctly. If performed incorrectly, the patient may not be able to see beneficial results. This establishes the need for an automated rehabilitation system that can guide the patients and help them carry out the exercises correctly, while also reducing the work of a doctor.

EMG signals have been used in the past to diagnose neuromuscular disorders, for controlling robots using hand gestures, and so on. The placement of EMG electrodes is also important to ensure that proper readings are taken. It is advised to increase the density of the EMG leads in relation to the circumference of the EMG sensor system’s circular form to obtain more accurate results [3]. In the field of athletics, EMG signals were used to detect the first and second breakpoints, that is, the aerobic and anaerobic thresholds [4]. When it comes to activities like running, EMG signals have been used to assess the inertia and force exerted by the muscles to evaluate sports performance [5].

EMG has been extensively used in evaluating the extent of rehabilitation. EMG has been used in combination with electrocardiogram (ECG) signals of the heart in order to analyze upper limb rehabilitation [6]. The upper limb movements are analyzed using EMG, whereas the heart’s status is assessed using ECG signals. Many home-based rehabilitation devices have been developed that make use of EMG signals. Smart wearable devices that acquire EMG signals and process them to assess the user’s muscle function from their homes have been developed [7]. The pathophysiology of a muscle was analyzed using EMG signals for the rehabilitation of cerebral palsy [8]. The speed with which the muscle fiber cells shoot electrical impulses is called muscle fiber conduction velocity (MFCV) [9]. Studies have shown that EMG signals are a reliable tool for investigating MFCV [10]. EMG signals can also be used in combination with electroencephalography (EEG) signals as a support signal in neuromotor rehabilitation [11]. Various signal processing techniques have been developed for analyzing EMG signals, in addition to the ones that already exist. Wavelet analysis has been performed on muscles to quantify muscle coactivation of patients with spinal cord injury [12].

It is now evident that EMG signals have been used successfully for rehabilitation in the past. In this study, EMG signals have been used exclusively to evaluate the correctness of exercises carried out by DRA patients. The primary objective of this research is to evaluate the correctness of core-strengthening, abdominal exercises in the rehabilitation of postpartum patients experiencing DRA. To the best of our knowledge, this is the first work reporting the objective evaluation of rehabilitation exercises of DRA. It is of utmost importance to ensure that these exercises are being carried out correctly while following all necessary procedures. The movement-related details were captured with the help of wearable EMG sensors. In addition to the traditional time and frequency domain features, certain non-linear features were also analyzed to select the optimal feature set for classification. Different machine learning models were trained with these features through a series of experiments. Ensemble learning methods involving RUSBoosting provided the best performance when compared to other models.

The rest of the manuscript is organized as follows. Section 2 presents an overview of the proposed methodology. Details of the different experiments performed and their analyses were covered in Section 3. Section 4 presents the concluding remarks of this research work along with the scope for future work.

2. Proposed Methodology

The workflow of the proposed research work is presented in Figure 1.

2.1. Muscle Anatomy of DRA

The rectus abdominis muscle is a paired stretch of muscles located in the abdominal region of humans. This parallel set of muscles is separated from each other by a connective tissue forming a midline band. This band is called the linea alba [13]. The width of the linea alba determines the distance between the two recti muscles that are parallel to each other. These recti muscles are typically 10 mm thick, although they can be thicker (~20 mm) in athletes and bodybuilders [14]. There are certain tendinous intersections that further divide the recti muscles into smaller muscle groups. The top part of the recti muscles is primarily attached to the fifth rib. It is evident from Figure 2 that in a normally functioning abdominal muscle, the width of the linea alba is very small. These muscles play a very important role in maintaining proper posture of the upper body owing to their proximity to the pelvic region and the lumbar spine. During pregnancy, the uterus expands to accommodate the fetus. This results in major strain on the recti muscles, causing the connective tissue (linea alba) to separate. However, these muscles should return to normalcy post-pregnancy. That is, the width of the linea alba should be reduced. When that does not happen within the first eight weeks postpartum, it results in a condition called Diastasis Recti Abdominis, where the abdominal muscles remain separated, as seen in Figure 2. It may first present itself as a ridge running vertically across the midline of the abdomen starting anywhere from the xiphisternum (the breastbone) to the umbilicus. This state of weakened abdominal musculature often causes major discomfort to those experiencing it, such as lower back pain, pelvic pain, difficulty lifting objects, and improper posture. One of the most common physical symptoms is the presence of a prominent bulge in the abdomen’s midsection. In some cases, a hernia might develop as a complication. It is diagnosed through physical examination or through an abdominal ultrasonography. First-stage treatment is provided through physical exercises that work towards contracting the recti muscles towards each other, in order to reduce the gap. In some very extreme cases of DRA, a cosmetic surgery of the abdomen called abdominoplasty is carried out to correct the muscles [15].

2.2. Role of Therapeutic Exercises

As discussed previously, first-stage treatment for DRA involves physical exercises that aim to contract the recti muscles to bring them closer to each other. Previous studies have shown that such core strengthening exercises may or may not help in bringing the recti muscles together. Nonetheless, these exercises are only beneficial if performed correctly since the chances of a cure are better in this case. Incorrectly performed exercises may increase the separation of the recti muscles, thereby worsening the condition. Crunches are exercises that increase the separation of these muscles and should be avoided at all costs. All corrective exercises should focus on bringing the muscles together and should avoid pushing them apart [16]. A description of the exercises prescribed in this study are as follows:

• SIT-UPS (SU)

Sit-ups are a very popular core-strengthening exercise that tighten the abdominal muscles. It begins with lying on one’s back on the floor, usually with arms across the chest or palms behind the head and the knees and toes bent (to reduce stress on the back muscles and spine), and then elevating both the upper and lower vertebrae from the floor as seen in Figure 3.

2.

CURL-UPS (CU)

It is important to ensure that the knees are bent and that the feet are rested comfortably on the floor. One should adjust the pelvic region to ensure that the lumbar spine is in a neutral position. Keeping the neck fixed, the head is curled thereby bringing the chin towards the chest. The arms, legs and shoulders should be relaxed. Only the upper body should move with the focus on the abdomen. If the movement is performed correctly, the head and arms will curl up as one unit with the shoulders.

3.

TRANSVERSE ABDOMINIS WITH SIT-UPS (TASU)

One should first lie down on their back comfortably, with their spine in a neutral position. The pelvic floor is to be contracted by drawing the muscles from behind the pubic bone to the tailbone, and the left and right sides of the pelvic floor together. Then one should gently draw the entire pelvic floor up, similar to zipping a drawstring bag. This is a transverse abdominis (TA) contraction. Once the muscle is in that contracted state, a sit-up should be performed.

4.

TRANSVERSE ABDOMINIS WITH CURL-UPS (TACU)

One should first lie down on their back comfortably, with their spin in a neutral position. The pelvic floor is to be contracted by drawing the muscles from behind the pubic bone to the tailbone, and the left and right sides of the pelvic floor together. Then, one should gently draw the entire pelvic floor up, similar to zipping a drawstring bag (TA contraction). Once the muscle is in that contracted state, a curl-up should be performed, as shown in Figure 4.

5.

TRANSVERSE ABDOMINIS, PELVIC FLOOR MUSCLES AND CURL-UP (TAPF-CU)

To perform this exercise, one should lie down on one’s back and perform a TA contraction. The most important part of this exercise is to take a deep breath and allow the stomach muscles to rise as it fills with air. The pelvic floor muscles should be kept relaxed as the breath is taken in. Exhalation should now happen through the mouth while gently contracting the pelvic muscles. The pelvic floor muscles should be kept contracted for about 3 to 6 s, until the muscles begin to get tired. This is called pelvic floor muscle contraction (PFM contraction). Upon the next inhalation, the contraction is released, and the muscles are relaxed. The muscles should relax for about six to ten seconds. Along with the TA contraction and a PFM contraction, a curl-up is to be performed.

6.

STRAIGHT LEG-RAISE (SLR)

To perform this exercise, one should lie down with hips squared and legs laid flat on the floor. One knee is to be bent at a 90-degree angle while the foot is flat on the floor. The thigh should be contracted to stabilize the muscles on the straight leg. Inhalation should be performed slowly, the straight leg should be raised about six inches from the ground, and this is to be maintained for about 3 s. Exhalation should happen slowly and the raised leg should be lowered to the floor with control.

7.

REVERSE CURL-UP (RCU)

To begin this exercise, one should lie down on one’s back with one’s arms by the side. Both legs should be raised such that the thighs are perpendicular to the floor, while the knees are bent at a 90-degree angle. This is followed by exhalation and a contraction of the abs to bring the knees up towards the chest. A slight rise of the hips is needed for a beat. Slowly, the legs are to be brought back to the starting position.

2.3. Data Acquisition and Experimental Setup

It can be seen from Figure 1 that the first step is to acquire good quality EMG signals. The EMG signals are acquired using Shimmer3 (F087 and F172) EMG recording systems that come with disposable surface EMG electrodes. The sensors are placed on the target muscles in the abdomen and signals are acquired from two different channels, Channel 1 and Channel 2.

Table 1. Specifications of the EMG recording unit used.

Feature	Specifications
Input protection	Radio Frequency and Electromagnetic Interference filter (3 MHz) Defibrillation protection
Analog-to-Digital Convertor (ADC)	24-bit ADC
Firmware for wireless transmission	LogAndStream v0.6.0
Transmission bandwidth	8.4 kHz
Sampling frequency	512 Hz
Weight	31 g
Dimensions	63 mm × 32 mm × 12 mm
EEPROM memory	2048 bytes
Ground	Wilson Type Driven Ground
Gain (configurable)	1, 2, 3, 4, 6, 8, 12
Input differential dynamic range (for gain 6)	800 mV
Connections (Input sensor electrodes)	Channel 1 positive—Ch1P Channel 1 negative—Ch1N Channel 2 positive—Ch2P Channel 2 negative—Ch2N Reference electrode—Ref

describes the hardware specifications of the EMG unit and Figure 5 represents its internal system architecture.

The electrodes of the sensors are placed on the surface of the skin. Hence, the acquired EMG signals are called Surface EMG signals. This is a non-invasive technique. Intramuscular EMG signals, however, are invasive in nature and were not used in this study. Figure 6 shows the Shimmer3 surface EMG recording unit. The placement of the sensor electrodes and the Shimmer3 unit on a subject who was about to perform an exercise is shown in Figure 7.

Datasets were acquired from twenty different subjects, each performing the seven different exercises (discussed earlier), both correctly and incorrectly. Figure 8 represents a sample raw unprocessed EMG signal when sit-ups were performed three times correctly.

Raw EMG signals, or any biological signal for that matter, contain a lot of noise from many sources [17]. Electrical line noise and power line interference are the most common sources of noise. In addition to that, noise may also be caused by the friction of electrodes against the skin. In some cases, temperature may also influence the readings taken. If the surface of the skin is not cleaned before signal acquisition, it may induce errors in the signals thereby resulting in erratic fluctuations. If the sensors are not calibrated properly, these may also introduce random and erratic movements in the amplitude of the signal. These noisy components must be removed before proceeding to feature extraction. This is carried out because random high spikes in the amplitude of the EMG signal would affect the results of analyses that take amplitude as in input.

2.4. Preprocessing of EMG Signals

The workflow of the preprocessing is presented in Figure 9.

Outlier Removal: The bandpass filtered signal taken from the output of Shimmer F087/F172 unit was then passed through for outlier removal. This was performed based on a standard deviation threshold. Any point of the signal that lies above the threshold was removed. These erratic points of the signal are usually random spikes in the signal caused by sensor malfunctions or motion artifacts. Figure 10 represents this process graphically.

Rectification and Smoothing: The amplitude points of an EMG signal tend to have both positive and negative values. The full-wave rectification of EMG signals is carried out to ensure that the signal points are all positive. Therefore, the extracted features do not result in a value of zero in analyses involving a summation of the signal points, such as the mean [17]. The absolute values alone are considered for the computation of such features.

Smoothing is performed to produce a smooth envelope of the overall signal. A moving average filter with thirty-five signal points was used for this step. It produces a signal that is an average of thirty-five points before and after each signal point under analysis. The pre-processed signal after smoothing is presented in Figure 11.

2.5. Feature Extraction

The processed signals were taken to the next stage: feature extraction. A range of features were extracted from the signals across different domains, namely, the time domain, frequency domain and the time–frequency domain, along with a few nonlinear features. A brief description of all the features is as follows:

1.

Energy

One of the most essential and basic features while processing signals is energy. It is the area under the squared magnitude of the signal over time [18].

$$E_s= {{\displaystyle \int }}_{-\infty }^{\infty }{\left|x\left(t\right)\right|}^2 dt$$

(1)

where Es is the energy of the signal and x(t) is the signal of interest in Equation (1).

The greater the amount of energy a signal has, the more accurate its recovery from noise is. With respect to EMG signals, the motive for extracting the energy is to ascertain whether one type of signal (correct or incorrect) has more or lesser energy than the other to enable differentiation.

2.

Kurtosis

Kurtosis is a statistical parameter. It gives one a measure of the number of peaks in a random signal. Higher the value, greater the number of peaks in the signal [19]. It is a dimensionless parameter. It tells us about the distribution of a signal relative to a Gaussian distribution. In a way, it tells us how outlier-prone the signal is in comparison to a normal distribution, which has a kurtosis value of 3. With respect to EMG signals, the goal was to analyze and see if the correctly performed exercises exhibited higher or lower kurtosis values than the other.

$$k=\frac{E{\left(x-\mu \right)}^4}{{\sigma }^4}$$

(2)

where k is the kurtosis, E(t) is the expectation of the value t, x is the distribution (here, signal), μ is the mean of x, and σ is the standard deviation of x in Equation (2).

3.

Power

Power of a signal is the sum of the squares of the amplitude points in the signal divided by the length of the signal. In place of a signal’s amplitude, the RMS of a signal may also be used in some cases.

$$P_x=\frac{1}{2T}{{\displaystyle \int }}_{-T}^T{\left|x\left(t\right)\right|}^{2}dt$$

(3)

where P_x is the power, T is the time period, and x(t) is the signal of interest in Equation (3).

The goal here was to uncover a possible existence of significant differences in power between the two types of signals in question.

4.

Mean frequency

This feature is taken from the frequency domain of a signal. First, a time domain signal’s power spectrum is calculated. Then the mean-normalized frequency of the power spectrum is computed. This gives us the mean frequency [20].

$$S_{xx}\left(f\right)=E\left[\frac{{\left|X_T\left(f\right)\right|}^2}{2T}\right]$$

(4)

where S_xx(f) is the power spectral density, X is the signal, and T is the time period in Equation (4).

The signal is first passed through the above formula for the computation of the power spectral density and then its mean-normalized frequency is calculated.

5.

Median frequency

This feature is taken from the frequency domain of a signal. First, a time domain signal’s power spectrum is calculated. Then the median-normalized frequency of the power spectrum is computed. The signal is first passed through the above formula for the computation of power spectral density and then its median-normalized frequency is calculated.

6.

Discrete Wavelet Transform

Discrete wavelet transform (DWT) has been used as a feature extraction technique for biomedical signals in many instances [21]. Daubechies wavelets are a family of orthogonal wavelets [22]. They too define a DWT. For a given support, they are defined or characterized by the number of vanishing moments. For this project, DWT was performed using 1st order Daubechies wavelets, and its approximation was taken at the 10th level. Two features were calculated from the 10th level of approximation, namely, D_med and D_rms.

D_med: This is the median of the approximated signal acquired at the 10th level of wavelet decomposition.

D_rms: This the root mean square value of the approximated signal acquired at the 10th level of wavelet decomposition.

7.

Entropy

Approximate entropy tells us about whether or not the signal is regular, that is, it gives us information about the regularity of the data [23]. The value of this entropy is high when the signal is highly complex or highly irregular. A regular and predictable time-series signal would typically have a very low approximate entropy.

$$N_i={{\displaystyle \sum }}_{i=1,i\ne k}^{N}1(\left||Y_i-Y_k\right|{|}_{\infty }<R$$

(5)

where Y is a reconstruction signal and R is radius of similarity in Equation (5). Then, entropy is calculated as in Equation (6).

$${\phi }_m={\left(N-m+1\right)}^{-1}{{\displaystyle \sum }}_{i=1}^{N-m+1}\mathrm{loglog} \left(N_i\right)$$

(6)

8.

Detrended Fluctuation Analysis (DFA)

This nonlinear analysis gives us the Hurst exponent as a result. This component tells us about the degree of self-similarity in a signal. That is, it tells us if a latter part of the signal resembles or at the very least, has some similarity with the earlier parts of a signal [24]. The Hurst component is given by:

$$H=\frac{\mathrm{loglog} \left(\frac{R}{S}\right) }{\mathrm{loglog} \left(T\right) }$$

(7)

where T is the time period duration, S is the standard deviation of the signal, and R is the difference between maximum and minimum deviation from the mean in Equation (7).

The (R/S) component is plotted against the T parameter in the logarithm domain. The slope of this resulting straight line gives us the Hurst exponent.

9.

Largest Lyapunov’s Exponent

This is a nonlinear feature that talks about the degree or convergence or divergence of a signal. Negative values indicate convergence, whereas large positive ones indicate divergence [25]. Values close to 0 indicate that the signal points maintain relatively similar phase-space points throughout the signal duration. The relation for the largest Lyapunov’s exponent is presented in Equation (8).

$$y\left(i\right)=\frac{1}{\Delta t}〈lnln d_j\left(i\right)〉$$

(8)

where d is the distance between each phase-space point and its nearest neighbor.

The summary of the extracted features is presented in

Table 2. Summary of extracted features.

Feature	Description	Inference
Energy	The area under the squared magnitude of signal.	Tells us about how much energy is contained within the signal. Goal was to see if correctly performed exercises had more/less energy.
Kurtosis	Kurtosis is a measure of how outlier-prone a distribution is. The kurtosis of the normal distribution is 3.	In spite of outlier detection and removal, some outliers are inevitable in analysis. The goal was to exploit that to see if this could be a suitable feature for classification.
Power	Maximum power of signal is related to spectrum analysis and is calculated using STFT.	It tells us about the maximum power contained in a correct/incorrect EMG signal and helps in analysis.
Mean Frequency	The average value of the normalized frequency taken in the frequency domain after computing power spectral density (PSD).	Mean and median frequencies are useful in estimating muscle fatigue.
Median Frequency	The median value of the normalized frequency taken in the frequency domain after computing the power spectral density (PSD).	Mean and median frequencies are useful in estimating muscle fatigue.
D_median	Median of detailed coefficients of the 10th level, 1st order Daubechies wavelet decomposition	WT reduces crosstalk between different muscles. In the 10th level of WT and decomposition, it tells us if the median of the EMG signal shows characteristics that can differentiate between correct and incorrect.
D_rms	RMS of detailed coefficients of the 10th level, 1st order Daubechies wavelet decomposition	WT reduces crosstalk between different muscles. In the 10th level of WT and decomposition, it tells us if the RMS of the EMG signal shows characteristics that can differentiate between correct and incorrect.
Approximate Entropy	Approximate entropy is a regularity statistic that quantifies the unpredictability of fluctuations in a time series. High value indicates that similar patterns are not followed by the same.	EMG signals are nonstationary signals that vary erratically with time. Such a time series signal is prone to fluctuations. Here the goal is to see whether correct and incorrect exercises produce differentiable EMG in terms of predictability.
DFA—Hurst exponent	A nonlinear feature that tells us about the degree of self-affinity in a signal.	EMG signals are nonstationary signals, and their characteristics vary with time. DFA tells us about the predictability of the latter part of a signal in comparison to its initial parts.
Largest Lyapunov’s Exponent	It quantifies the rate of divergence or convergence of close trajectories in phase space. positive indicates divergence, negative indicates convergence.	In the phase space, the goal was to analyze whether or not correct and incorrect exercises produce convergence/divergence.

.

2.6. Feature Selection

While modeling a classifier, it is essential to ensure that only the necessary features are taken into account. Making the classifier learn unnecessary features that do not influence the response variable might result in overfitting. The chi-square test and minimum redundancy maximum relevance (MRMR) algorithms are used in this study to perform feature selection [26,27]. Common features predicted by these algorithms along with other combinational trials will be used as inputs to the classifier.

2.7. Classification

The selected features will be given as inputs to the classifier. The classifier will be trained to make predictions using the processed EMG signals, in order to determine whether or not the exercises are being performed correctly. Trials are carried out for various ensemble learning algorithms in order to analyze all possible combinations and pick the best performing one. Since the classification process is carried out on labeled data, it falls under the purview of supervised machine learning [28]. In this study, the results belong only to two groups, namely, correctly performed exercises and incorrectly performed exercises; hence, the model would act as a binary classifier. The data was split into training and testing datasets before being fed into the model. For this study, the train–test split was 90–10%. Training was performed with five-fold stratified cross validation. Validation was carried out to tune the hyperparameters and train the model to increase overall accuracy. Ensemble learning was chosen for this study owing to the fact that it is more robust than single decision trees. They combine multiple decision tree classifiers to make a predictive model [29]. Bagging and boosting are two types of ensemble learning algorithms. Bootstrap aggregation, commonly referred to as bagging, works by learning the training data in multiple subsets and taking an overall average of each subset’s accuracy. It also decreases the variance in the predictions it makes and prevents overfitting. Multiple subsets are created from the original data. Each of these subsets is trained in parallel and independent of each other. The final predictions are made by combining the predictions from all these models [30]. The random forest algorithm uses bagging to make predictions. Boosting is based on progressive improvement of the base model. Once an initial prediction model is developed, a second one is developed that tries to correct the errors of the first. This process continues either until all values are predicted correctly or until the number of models has been specified [30]. In this study, an exploration of different bagging and boosting algorithms was carried out, including RUSBoost, AdaBoost, LogitBoost, and GentleBoost.

AdaBoost is an ensemble learning algorithm that trains learners sequentially. For each learner, the AdaBoost algorithm computes the weighted classification error. The word ‘adaptive’ is used due to the fact that this algorithm classifies by increasing the weight of the observations that were initially misclassified by the learner and also simultaneously reduces the weight of those classified correctly. The subsequent learner will make use of the updated weights to classify [31]. In this study, RUS Boosting, that is, random undersampling boosting was used for binary classification owing to the fact that the data was imbalanced in nature. When one class seems to occur more than the other in the case of binary classification, the data is said to be imbalanced. This algorithm first takes n members of the class with the lower number of occurrences for training. The class with more members is undersampled with only n observations [32]. Logit Boost, that is, adaptive logistic regression, has a working method similar to AdaBoost but it minimizes binomial deviance. This algorithm works better on data with poorly separable classes. Here, every weak learner fits a logistics regression model to the response values for making predictions [33]. GentleBoost is a combination of both AdaBoosting and LogitBoost. It minimizes exponential loss while also fitting a regression model to make predictions.

3. Results and Discussion

This section compiles the results acquired from the processes carried out.

3.1. Graphical Analysis of Features

This section graphically represents the box plots of the three top performing features with the best differentiating capabilities. The results displayed in

Table 3. Box Plots of the three top features.

Feature	Box Plots
Kurtosis
Mean Frequency
DFA—Hurst Exponent

are the results of initial exploratory data analysis.

Exploratory data analysis has revealed that the features in Table 3 seem to have good chances of being suitable features to give as input to the classifier. These features, along with other suitable ones deduced through the feature selection process have high chances of yielding good results. From Table 3, it can be said that the values of kurtosis (the degree of how outlier-prone a signal is) and mean frequency seem to possess a range within which the values of correct exercises and incorrect exercises seem to lie (based on visual inspection). DFA (giving the Hurst exponent as a result), does not have such a well-defined range. However, the values of DFA, when used in combination with other features, may be useful in performing the classification of EMG signals.

3.2. Analysis of Feature Selection

A total of ten features were extracted across different domains. Given that the goal is to determine whether or not exercises are being performed correctly or incorrectly, the best possible features of all those extracted need to be selected. These features would be most suitable for classification. For this very crucial feature selection process, the chi-square test and the minimum redundancy maximum relevance (MRMR) algorithms were used.

Chi-square tests are a filter-type feature selection algorithm. They are used to rank the features based on their relationship with the response variable. The predictor variables are subject to individual chi-square tests and their influence on the response variable is evaluated. A small p-value as a result of the test indicates that that particular predictor variable is an important one for classification. The acquired ten features were stored as a dataframe and were fed as input into this statistical feature selection algorithm.

The MRMR algorithm is also a filter-type feature selection algorithm that selects highly predictive features that are least correlated with each other. i.e., it selects the features that have the highest correlation with a class (relevance) but the least correlation between themselves (redundancy). Figure 12 and Figure 13 represent the ranking of features by the two feature selection algorithms.

Table 4. Top six features of both feature selection algorithms.

Chi-Square Algorithm	MRMR Algorithm
Kurtosis	Kurtosis
Mean frequency	Exercise performed
DFA (Hurst Exponent)	Mean frequency
Entropy	D_rms
D_med	Median frequency
D_rms	Entropy

tabulates the top six features from both these algorithms.

The common features between the two algorithms are: kurtosis, mean frequency, entropy, and D_rms. It was found that while MRMR predicts the features that are least redundant and least correlated with each other, chi-square tests evaluate the importance of each feature with the response variable. In multiple combinational trials, it was found that the features ranked by the chi-square test when used in combination with median frequency (from MRMR) in place of D_rms were more efficient than those ranked solely by the MRMR algorithm. As a result, the final top six features taken for classification are:

• Kurtosis
• Mean frequency
• DFA (Hurst exponent)
• Entropy
• D_med
• Median Frequency

3.3. Classification Model

The behavior of the EMG signals is different for correct and incorrect exercises. Based on visual inspection of the signals, it was seen that correctly performed exercises follow a more refined pattern of contraction and relaxation. A more erratic behavior with no clearly discernible contraction or relaxation is seen in the case of incorrectly performed exercises. This type of behavior is quantitatively represented by the extracted features. The model then learns this behavior of features to classify correctly done and incorrectly performed exercises. RUSBoosting is an ensemble learning algorithm that follows the method of decision trees to perform the classification. There is no single feature that stands out for classification. The behavior of all these features in combination with each other is capable of performing the differentiation.

Features were extracted from different domains. The best features that were most suitable for predicting the class (correct/incorrect) of an exercise using processed EMG signals were selected. The classification model takes into account the selected features and analyses those features to produce results. The result in this case, would be to ascertain the correctness of an exercise performed. Five-fold stratified cross-validation was performed on the dataset. The best performing model was the RUSBoosted algorithm with a validation accuracy of 88.4%. The confusion matrix obtained for this model is presented in Figure 14.

Table 5. Comparison of different classification models.

Model	Parameters	Training Accuracy (%)	Testing Accuracy (%)
Bagged Trees	Maximum number of splits: 20 Number of Learners: 5 Learning rate: 0.1	97	84.6
AdaBoost	Maximum number of splits: 20 Number of Learners: 30 Learning rate: 0.1	60.1	53.8
RUSBoosted	Maximum number of splits: 20 Number of Learners: 30 Learning rate: 0.1	100	76.9
RUSBoosted	Maximum number of splits: 35 Number of Learners: 35 Learning rate: 0.01	98.47	92.3
LogitBoost	Maximum number of splits: 20 Number of Learners: 25 Learning rate: 0.01	94.9	84.6
GentleBoost	Maximum number of splits: 20 Number of Learners: 30 Learning rate: 0.1	100	92.3

presents the comparison of classification accuracies of some well-known classifier models.

While both the RUSBoosted algorithm and the GentleBoost algorithm seemed to provide the same test accuracy, the former is preferred, owing to its reduced computational complexity and enhanced classification speed. It classified at the rate of 930 observations per second (3.1548 s in total), whereas the GentleBoost algorithm did the same at 750 observations per second (3.5329 s in total).

3.4. Discussion

The most important takeaway from the analyses carried out is that it is possible to differentiate between correctly performed exercises and incorrectly performed exercises by processing surface EMG signals. While not all classifiers were able to carry out said classification, some algorithms were capable of doing so. Models can be further developed by increasing the amount of training data and by tuning the hyperparameters.

EMG signals were acquired from two different channels, namely, Channel 1 and Channel 2. Signals acquired from either channel were seen to behave similarly. Each feature extracted had a possible purpose to serve, that is, a factor that might help with differentiation. However, it can be seen from the eventually selected features that no feature that involved computations using time-series amplitude was selected. This is due to the fact that amplitude changes and fluctuations vary from person to person, depending on the strength of electrical impulses generated by the muscle fiber cells. Even though such erratic sharp fluctuations were removed during the outlier removal stage of preprocessing, there is no standard or uniformity that one can expect from an EMG signal (taken from different people) with respect to its amplitude. Features extracted from the time domain were not very effective, whereas those taken from the other domains proved to be useful for classification. A combination of two different feature selection algorithms was chosen. Kurtosis as a feature discusses how outlier prone a signal is. During the preprocessing stage, all erratic outliers were removed. Nonetheless, kurtosis was computed to exploit the outliers that were ignored during preprocessing. This provided fruitful results, considering the fact that kurtosis was a highly ranked feature for classification. With respect to frequency domain features, mean and median frequencies were computed. These features were not only useful for classification but also for assessing muscle fatigue. Nonlinear features, such as DFA and entropy, proved to be useful features for classification. This is expected since the nonlinear and nonstationary nature of EMG signals are best explained by nonlinear features.

With respect to the classifier, various types of ensemble learning models were trained. The hyperparameters were tuned to increase the accuracy of the model. From Table 3, it is clear that the RUSBoosted algorithm with its various subtypes and tuned hyperparameters performed best for the data in this study. It cannot be said that all Boosting algorithms will work effectively. AdaBoost, although a very powerful algorithm, does not seem to work well with respect to the EMG data in this study. This could be due to the fact that biosignals are noisy in nature and some outliers might work their way into the data, thereby reducing the efficiency of an AdaBoost model. The RUSBoost algorithm worked much better when the learning rate was reduced from 0.1 to 0.01 and the number of learners was increased from thirty to forty. This algorithm clearly works well for imbalanced datasets and can be made to learn more efficiently when the hyperparameters are tuned. In comparison to the GentleBoost algorithm, the RUSBoost algorithm at a learning rate of 0.01 is preferred. The GentleBoost algorithm was more computationally intensive and took a longer time to perform the classifications. The RUSBoosted (learning rate of 0.01) algorithm provided results of the same quality and took less time to do so.

4. Conclusions

The main objective of this work was to study and analyze the correctness of core-strengthening exercises that target the abdominal muscles. These rehabilitation exercises are prescribed by the physiotherapists for DRA patients. The goal of these exercises is to bring the recti muscles closer to each other, thereby reducing the inter-recti distance. It is of utmost importance to ensure that these exercises are performed correctly for the faster resolution of DRA. The data regarding the movement of the recti muscles were acquired using Shimmer3 wireless EMG sensors. The input signals were pre-processed, and features were extracted using differing techniques. In addition to the traditional time and frequency domain features, certain nonlinear features were also employed. These nonlinear features were observed to explain the random nature of biological signals effectively. The best set of discriminating features were then used to train multiple machine learning models through a series of experiments. RUSBoosting, an ensemble learning algorithm provided the best performance when compared to the other models. This research can further be extended with a larger group of subjects using deep learning algorithms.