Assessing aberrant muscle activity patterns via the analysis of surface EMG data collected during a functional evaluation

The data had been previously gathered using two EFA-based protocols consisting of a battery of static and dynamic tests designed to assess individuals with cervical pain and back pain, respectively. The dataset included recordings from 62 subjects of age ranging between 22 and 61 years. Fifty-six of these individuals were males. Forty-six of the subjects were evaluated using both protocols. The remaining 16 subjects were evaluated using only one protocol.

For the first protocol, surface EMG recordings were gathered bilaterally from the following muscles of the neck, shoulders, arms, and the thoracic section of the back: scalene, upper trapezius (two channels were recorded using electrodes positioned on the descending muscle fibers below the occipital bone and on the transverse muscle fibers near the acromion, respectively), middle trapezius, longissimus thoracis, middle deltoid, biceps brachii, and triceps. Surface EMG recordings were collected during two static tests, namely 1) while subjects sat still for about 20 s (a condition herein referred to as rest sitting) and 2) while they stood without moving for about 20 s (a condition herein referred to as rest standing). Surface EMG recordings were also collected during three dynamic tests: 1) flexion/extension, rotation, and lateral movement of the head (herein referred to as head movements), 2) shoulder shrug and overhead reaching with both arms (herein referred to as shoulder and arm movements), and 3) a lifting functional capacity evaluation (FCE) (i.e. pulling a bar attached to a load-cell with the bar positioned at waist level and the knees fully extended) first with pronation of the forearms and then with supination of the forearms (this condition is herein referred to as FCE lifting with extended knees).

For the second protocol, surface EMG data was gathered bilaterally from the following muscles of the lower back and legs: paraspinal at L2, latissimus dorsi, gluteus maximus, biceps femoris, rectus femoris, tibialis anterior, and gastrocnemius (medial head). Surface EMG recordings were collected during the same static tests as per the protocol summarized above (i.e. rest sitting and rest standing) and during three dynamic tests: 1) flexion/extension, rotation, and lateral movement of the trunk (herein referred to as trunk movements), 2) taking two steps, kneeling, and standing up (herein referred to as walking and kneeling), and 3) a lifting functional capacity evaluation (i.e. pulling a bar attached to a load-cell with the bar positioned at knee height, the knees in a flexed position, and the arms fully extended) first with pronation of the forearms and then with supination of the forearms (this condition is herein referred to as FCE lifting with flexed knees).

Disposable adhesive silver/silver-chloride electrodes were utilized to collect the EMG data. The electrodes were circular in shape (approximately 5 mm in diameter) and had a layer of conductive gel to achieve low skin-electrode impedance. The Surface Electromyography for Non-Invasive Assessment of Muscles (SENIAM) recommendations [43] were followed to position the electrodes on the muscle belly of the selected muscles. The distance between the centers of electrode pairs was approximately 15 mm. Before attaching the electrodes to the skin, the skin was carefully cleaned using alcohol wipes and shaved when necessary. The electrodes were connected to wireless Bluetooth units (Shimmer Research, Dublin Ireland) that transmitted the EMG data to a base station connected to a computer. The wireless units were equipped with a DC-coupled unit (ADS1292R by Texas Instruments, Dallas TX) that provided the analog front-end as well as the analog-to-digital conversion. The data was sampled at 1024 Hz. The data was available prior to initiation of the study as part of a database meant to be used to develop algorithms to facilitate the analysis of data collected during functional evaluations. The database also contained data collected using a load-cell (LSB302 by Futek, Irvine CA) that relied on the above-mentioned wireless Bluetooth units and base station to simultaneously collect EMG and force data when required by the EFA-based protocol (Emerge Diagnostics, Carlsbad CA).

An EMG expert had previously annotated the surface EMG recordings performed during the above-described EFA-based protocols [34‐36] using criteria commonly used in a clinical context to capture the level of muscle activity (including its laterality), the presence/absence and severity of muscle spasms, and the muscles’ motor unit recruitment characteristics associated with the EMG amplitude modulation. Clinical scores were generated via visual inspection of the EMG data using an ordinal scale from 0 to 10 designed to capture the magnitude of each phenomenon of interest (e.g., an activity level score equal to 0 was associated with recordings marked by very low amplitude of the EMG data, whereas an activity level score equal to 10 was associated with recordings marked by very large ampitude of the EMG data). Surface EMG recordings gathered during the static tests were associated with clinical scores for activity level and spasm severity. Surface EMG recordings gathered during the dynamic tests were associated with clinical scores for activity level and amplitude modulation. From the clinical scores for activity level, the EMG expert also derived laterality of activity scores by computing the absolute value of the difference between the scores for the right and the left side of the body for each muscle.

Figure 1 shows an example of the EMG recordings selected from the dataset utilized in the study. The recordings shown in the figure were collected from the paraspinal muscles at L2 during the rest sitting test (Figs. 1a and b) and the FCE lifting with flexed knees test (Figs.1c and d). Figures 1a and c show data collected from a subject with very low level of muscle activity (activity level = 0) during the rest sitting test and high level of EMG activity and amplitude modulation (activity level = 10 and amplitude modulation = 10) during the FCE lifting with flexed knees test. Figures 1b and d show data collected from a subject displaying an elevated level of muscle activity (activity level = 5) during the rest sitting test and a modest level of activity and minimum amplitude modulation (activity level = 3 and amplitude modulation = 2) during the FCE lifting with flexed knees test.

A low level of muscle activity is expected in physiological conditions during the rest sitting test. In contrast, an elevated level of muscle activity is often observed in subjects with back pain. An elevated level of activity of the back muscles has been associated with postural changes that are common acutely in individuals who suffered a back injury and chronically in those who experience chronic back pain [26]. The recordings carried out during the static tests were also rated for spasm severity. In individuals with back pain, muscle spasms are believed to be triggered by the activity of nociceptive receptors responding to strain and soft tissue damage [17]. They can be observed in the EMG recordings as short bursts of activity or as sustained activity associated with large action potentials due to the synchronization of motor units [28]. The evaluation of the severity of muscle spasms in the latter case requires close inspection of the EMG recordings to identify the presence of action potentials at quasi-regular time intervals. Using these criteria, the EMG expert rated the recordings in Figs. 1a and b for spasm severity as 0 and 5, respectively. A high level of EMG activity and prominent amplitude modulation is expected when healthy subjects perform a functional capacity evaluation. The latter is expected because of the pattern of motor unit recruitment and de-recruitment that marks each burst of EMG activity associated with changes in the force generated by the muscle/s of interest. In contrast, a modest level of activity and minimum amplitude modulation is expected in individuals with back pain who display avoidance patterns [12].

Algorithms were developed using the Matlab (The MathWorks, Natick MA) programming environment. Estimates of the clinical scores were derived via the analysis of surface EMG data by filtering the raw EMG recordings, computing the EMG envelope, and extracting data features from the filtered EMG data and from the EMG envelope. Filtering the raw EMG data was achieved using an 8th order elliptic band-pass filter with cut-off frequencies equal to 20 and 400 Hz. This choice of cut-off frequencies allowed us to significantly attenuate low-frequency components associated with motion artifacts as well as high-frequency components associated with the instrumentation noise. The EMG envelope was derived by low-pass filtering the full-wave rectified raw EMG data with a 7th order elliptic filter with cut-off frequency equal to 10 Hz. We empirically determined the optimality of this cut-off frequency as suitable to preserve “high-frequency” amplitude modulation components relevant for an accurate estimation of clinical scores (e.g. spasm severity scores). Then different data features were derived to estimate each clinical score. Estimates of the activity level scores for the static tests were derived using the root-mean-square (RMS) value of the EMG envelope. This data feature appeared to be suitable to address the problem at hand because we observed that the EMG expert generated the activity level scores based on visual inspection of the EMG data amplitude. Estimates of the spasm severity scores for the static tests were derived using three data features: the RMS value of the filtered EMG data and the EMG data energy between 50 and 100 Hz and between 100 and 200 Hz. The energy of the EMG data in each of the above-mentioned bandwidths was estimated by deriving the periodogram of the EMG time series (using 1 s rectangular windows with 50% overlap) and integrating the EMG frequency components in the bandwidth of interest. These data features were selected because the EMG expert generated the spasm severity scores by considering both the amplitude of the EMG data and its frequency content. The latter was done in an attempt to identify when a prevalent number of type II fibers vs type I fibers were recruited. The frequency bandwidths utilized to derive the above-mentioned parameters were determined empirically. Finally, estimates of the clinical scores for the dynamic tests were derived using the following EMG data features: the RMS value of the EMG envelope, the range spanned by the filtered EMG data (i.e. the peak-to-peak amplitude of the time series), the RMS value of the EMG envelope during the time intervals when the muscle was silent or minimally active, the duration of time during which the muscle was active, the RMS values derived for the three main bursts of EMG activity, the average RMS value of the dominant frequency components of the EMG data, and the variance of the EMG envelope during the three main bursts of activity. The bursts of EMG activity were identified by using a 1 s sliding (by steps of 0.25 s) rectangular window, estimating the RMS value of the EMG envelope within each 1 s time interval, and selecting non-overlapping time intervals associated with the three largest RMS values. The average RMS value of the dominant components of the EMG data was derived by detecting the intervals during which the EMG envelope exceeded 1.5 times the RMS value of the EMG envelope during the entire test. These data features were selected because we observed that the EMG expert generated the clinical scores by considering the amplitude of the EMG data and characteristics of its modulation that appeared to be well captured by the above-mentioned parameters.

Estimates of the activity level scores for the static tests were derived by using the RMS value of the EMG envelope as a “proxy” of the clinical scores. This choice was motivated by the observation that the activity level scores and the corresponding RMS values of the EMG envelope appeared to be highly correlated. This choice was made based on a qualitative observation of the data and later confirmed by our quantitative analyses (see the Results section). Activity laterality scores were estimated by taking the absolute value of the difference between the activity level scores derived from the EMG recordings for the right and the left side of the body for a given muscle. The spasm severity scores were estimated by using a linear regression model relying on the above-mentioned EMG data features as independent variables. Finally, a linear regression model as well as a regression implementation of a Random Forest with the above-listed EMG data features as input variables were used to estimate the activity level scores and the amplitude modulation scores for the dynamic tests. A linear and a non-linear model for the analysis of the EMG data collected during the dynamic tests were implemented and compared because preliminary inspection of the EMG data features suggested that the output variables (i.e. the clinical scores) and the input variables were non-linearly related. The estimates of the activity level scores were also used to derive estimates of the activity laterality scores as described above for the static tests. Separate models were developed for the EMG data collected during each test.

The derivation of the linear regression coefficients for the static tests took into consideration the unbalance across classes corresponding to the different values of the clinical scores. This is a standard technique utilized in machine learning to avoid fitting well the data pertaining to classes with a large number of samples and fitting poorly the data pertaining to classes with a small number of samples [44]. Accordingly, because the instances of spasm severity scores equal to 0 were significantly larger than the instances for other score values, we randomly down-sampled to 10% of the total sample size the class associated with clinical scores equal to 0. We did so 10 times and hence derived 10 datasets to fit the model (i.e. 10 training sets). Also, because a relatively small number of instances with spasm severity scores greater than 5 was available, first, a model was developed using the instances with scores ranging from 0 to 5. Then, the model was applied to all the data. This process was repeated 10 times, for each of the datasets derived as mentioned above. Finally, the coefficient values of the model obtained for each of the 10 datasets were averaged. This approach was also used to derive linear regression models for the dynamic tests. The score distribution across classes for the dynamic tests was different than the score distribution for the static tests. The procedure was modified accordingly by down-sampling the class with the largest number of instances. The importance of each independent variable to generate the clinical score estimates was assessed based on the magnitude of the model’s standardized coefficients.

The Random Forest [37] models were derived using bootstrap aggregation (i.e. bagging [45]) of 100 regression trees with a minimum leaf size of 10. This method is very robust even when dealing with unbalanced classes. Hence, it was not necessary to rebalance the classes. A 10-fold cross-validation was used to derive the clinical score estimates. The importance of the input variables (i.e. the EMG data features) was assessed by measuring the increase in prediction error when the values of the input variables were permuted across the out-of-bag observations. This is a well-established method to assess the relevance of the input variables of a Random Forest [46].

The accuracy of the estimates derived using the above-described methods was assessed by deriving the root-mean-square error (RMSE) for each model. In addition, a linear regression between the clinical scores generated by the EMG expert and those estimated by the above-described EMG-based algorithms was computed for each clinical score. Regression coefficients as well as 95% confidence intervals were computed for each linear regression.

A subject with persistent back pain after instrumented lumbar fusion was evaluated using an EFA-based protocol prior to and after lumbar hardware removal. Before the lumbar hardware removal surgery, the subject reported low back pain with a severity level of 5 out of 10 on a visual-analog scale. The pain radiated to the left leg and was of higher intensity when stepping down with the left leg. Numbness and tingling were reported around the lumbar incision. The pain was clinically managed with medications. Inspection of the lumbar hardware during the surgery made apparent that one of the screws had migrated in the spinal canal and was likely to be responsible for the symptoms reported by the subject. The lumbar hardware removal led to a significant decrease in back pain, a decrease in pain medications, and an improvement in the subject’s functional ability.

The study was approved by the Western Institutional Review Board. The subject was evaluated using a modified version of the above-described protocol for the assessment of subjects with back pain. The battery of tests used to assess this subject included the following: 1) a series of rest sitting and a rest standing tests at the beginning of the evaluation; 2) a trunk flexion/extension movement test, which consisted of bending forward, returning to the upright position, bending backward, and returning to the upright position; 3) a trunk lateral movement test, which consisted of bending laterally first to the right and then to the left; 4) a walking and kneeling test; 5) an additional set of rest standing tests; 5) a series of FCE lifting with flexed knees tests; and 6) a final set of rest standing and rest sitting tests. During the evaluation, surface EMG data was collected bilaterally from the following muscles: paraspinals at L2 and L5, latissimus dorsi, gluteus maximus, biceps femoris, and rectus femoris. In addition, triaxial accelerometer recordings were gathered using wireless sensor units positioned at L1 and S1 to derive range of motion data. Also, a load-cell was used during the FCE tests.

Figure 2 shows a box plot representation of the results obtained using the above-described algorithms for the analysis of data collected during the performance of static tests. The plots show aggregate results for the rest sitting and the rest standing tests. For each class (i.e. for each clinical score value), the plots display the median value of the EMG-based estimates (red horizonal line), a box spanning the interquartile range of the distribution of the clinical score estimates, and whiskers that span the range from the minimum to the maximum estimated clinical score values. It is worth noticing that outliers (defined as values exceeding the first and third quartile more than 1.5 times the interquartile range) are not shown in Fig. 2. However, outliers were not removed when assessing the accuracy of the EMG-based estimates (i.e. when estimating the RMSE values as well as the regression coefficients and corresponding 95% confidence intervals). It is also worth noticing that when the number of instances in a class is small, the interquartile range and the range between the minimum and the maximum estimated clinical score values might be undistinguishable. Hence, the whiskers might be not clearly displayed in the box plot representation. This is the case for a few classes in Fig. 2.

Figure 2a shows the surface EMG-based estimates of the clinical scores for activity level. The estimates were derived using the RMS value of the EMG envelope as a proxy for the activity level scores. Although the plot shows aggregate results across testing conditions, separate models were derived for the rest sitting and the rest standing tests. The regression coefficients and corresponding 95% confidence intervals derived from the data collected during the rest sitting and the rest standing tests were equal to 0.89 ± 0.04 (regression coefficient ± 95% confidence interval value) and 1.19 ± 0.04, respectively. The small 95% confidence interval values associated with the regression coefficients for both the rest sitting and the rest standing scores are indicative of a good correlation between the EMG-based estimates and the clinical scores generated by the EMG expert.

Figure 2b shows the surface EMG-based estimates of the clinical scores for spasm severity. The estimates were derived using a linear regression model with the following three independent variables as input: the RMS value of the filtered raw EMG data and the EMG data energy between 50 and 100 Hz and between 100 and 200 Hz. Figure 2b shows aggregate results for the rest sitting and the rest standing tests. Regression coefficients and associated 95% confidence intervals were derived separately for the rest sitting (0.63 ± 0.02) and the rest standing (0.71 ± 0.02) tests. Both models used the above-mentioned independent variables as input variables. The standardized coefficients and corresponding 95% confidence intervals for the rest sitting test were 0.02 ± 0.07, 0.58 ± 0.05, and − 0.03 ± 0.06. The standardized coefficients and corresponding 95% confidence intervals for the rest standing test were 0.08 ± 0.11, 0.42 ± 0.07, and 0.14 ± 0.06. The values of the standardized coefficients suggest that the EMG data energy between 50 and 100 Hz is the most relevant independent variable to estimate the spasm severity scores.

Table 1 provides a summary of the above-discussed results. The table shows the RMSE values as well as the regression coefficients and associated 95% confidence interval values for the EMG-based estimates of the clinical scores for activity level and for spasm severity derived for the static tests. The table shows also the results for the laterality of activity scores. The RMSE values range from 0.37 (for the estimation of the spasm severity scores for the rest sitting test) to 0.98 (for the estimation of the laterality of activity scores for the rest standing test). The regression coefficient values range from 0.63 to 1.19 and the associated 95% confidence intervals range from 0.02 to 0.06. These results indicate that the proposed methods allow one to achieve accurate estimates of the clinical scores for activity level, spasm severity and laterality of activity scores.

Figure 3 shows an example of the surface EMG-based estimates of the clinical scores for the dynamic tests. Specifically, the figure shows the estimated amplitude modulation scores using a linear regression model (Fig. 3a) and using a Random Forest-based algorithm (Fig. 3b) vs. the scores generated by the EMG expert for the FCE lifting with flexed knees test. The box plot representation of the surface EMG-based score estimates shows that the Random Forest-based algorithm generated more accurate estimates than the linear regression model. This qualitative observation is confirmed by the estimation errors associated with these two different approaches to achieve the clinical score estimates. In fact, the RMSE of the surface EMG-based estimates shown in Fig. 3a (for the linear regression model) was 2.09 whereas the RMSE of the estimates shown in Fig. 3b (for the Random Forest-based model) was 1.74.

Tables 2 and 3 show a summary of the RMSE values as well as the regression coefficients and associated 95% confidence interval values for all the dynamic tests. Table 2 shows the values for the surface EMG-based estimates derived using linear regression models. Table 3 shows the values for the surface EMG-based estimates derived using Random Forest-based models. The tables show the RMSE values as well as the regression coefficients and associated 95% confidence interval values for all the test conditions for activity level, amplitude modulation, and laterality of activity. The RMSE values obtained using the Random Forest-based models were on average smaller than those obtained using the linear regression models with improvements approaching 50% for the FCE lifting with extended knees test. The regression coefficients and 95% confidence intervals values show a similar trend. The regression coefficients obtained using the linear regression models ranged from 0.84 to 1.72 with 95% confidence interval values ranging from 0.01 to 0.18. The regression coefficients obtained using the Random Forest-based models ranged from 0.88 to 1.04 with 95% confidence interval values ranging from 0.02 to 0.08. Overall, these results indicate that Random Forest-based models are more suitable than linear regression models to estimate activity level, amplitude modulation, and laterality of activity scores. Analysis of the relevance of the EMG data features used as input to the Random Forest-based models showed similar results for the activity level and amplitude modulation scores. In both cases, the most relevant input variables to estimate the clinical scores were the RMS values derived for the three main bursts of EMG activity, and the variance of the EMG envelope computed for the largest burst of muscle activity.

The improved accuracy of the EMG-based estimates obtained using Random Forest-based models over those obtained using linear regression models is apparent when one examines plots of the estimation errors as a function of the clinical scores generated by the EMG expert. An example of such plots is shown in Fig. 4. The plots in Fig. 4 allow one to compare the results obtained with the above-mentioned approaches when estimating the amplitude modulation scores for the FCE lifting with extended knees test. The plot on the left side (panel a) shows a clear trend in the bias affecting the EMG-based estimates. This is likely to be the case because the linear regression model fails to account for the non-linearity of the relationship between the input variables and the amplitude modulation scores. Such relationship is instead properly captured by the model based on a regression implementation of a Random Forest. In fact, the results show a modest trend in estimation bias and a relatively constant variance of the EMG-based estimates across clinical scores (panel b). Similar plots were obtained for all the EMG-based estimates for all the dynamic tests. Inspection of the plots showed similar differences as those shown in Fig. 4 between the estimates derived using a linear regression model and the estimates derived using a Random Forest-based algorithm. These observations indicate that Random Forest-based algorithms are more suitable than linear regression models to estimate clinical scores for recordings collected during dynamic tests.

Figure 5 shows an example of the surface EMG and load-cell data collected during the study. The panels of Fig. 5 show the EMG recordings from the left paraspinal muscle at L2 and the load-cell data collected during the performance of the FCE lifting with flexed knees test prior to (panels a and c) and after (panel b and d) the hardware removal surgery. The activity level and the amplitude modulation scores prior to the hardware removal surgery were 2 and 0, respectively. The corresponding clinical scores after the hardware removal surgery were 4 and 3, respectively. These clinical scores captured an increase in the EMG level of activity and its amplitude modulation. In fact, the RMS value of the bursts of EMG activity after the hardware removal surgery was approximately 40% larger than those detected before the surgery. The load-cell data also showed a marked increase in force output following the lumbar hardware removal surgery. In fact, the load-cell output barely reached 100 lbs before the surgery whereas it approached 150 lbs after the surgery. It is worth noticing that the improvements in the patterns of muscle activity shown in Fig. 5 were observed at the level of the lumbar fusion.

Additional observations of interest were gathered during the other tests performed as part of the EFA-based protocol that was used to clinically assess the subject. During the rest standing test at the beginning of the battery of tests, the analysis of the data identified a large level of activity in the right biceps femoris that decreased after the lumbar hardware removal surgery. Spasms were observed in the recordings from the left biceps femoris before surgery that were not present after surgery and compensatory activity in the left rectus femoris that was not present after surgery. The analysis of the surface EMG data collected during the trunk flexion/extension movement tests showed an improvement in the amplitude modulation of the data collected from the paraspinal muscles at L2 and at L5, and the gluteus maximus. The data collected during the trunk lateral movement tests showed improvements in the patterns of activity of the right and left latissimus dorsi muscles, the paraspinal muscles at L2 and at L5, and the biceps femoris. The data collected during the walking and kneeling tests showed bilateral improvements post-surgery in the patterns of activity of the paraspinal muscles at L2 and at L5. Improvements post-surgery were also observed bilaterally in the activity of the paraspinal muscles at L2 and at L5 during the rest standing tests that followed the walking and kneeling tests and in the activity of the biceps femoris and rectus femoris during the rest standing tests that followed the FCE lifting with flexed knees tests. No differences were observed during the rest sitting tests at the end of the battery of tests.

These observations are clinically relevant. The results of the above-summarized analyses indicate that the subject displayed patterns of muscle activity post-surgery that were closer than the ones observed before surgery to the physiological patterns of activity. These EMG-based observations were associated with improvements in range of motion, force output and in the general status of the individual (e.g. self-report of pain severity). Most importantly, close inspection of the hardware during the surgery showed that one of the screws had migrated in the spinal canal and was likely causing the symptoms reported by the subject. These results should be looked upon as a preliminary example of potential application of the procedures developed in the study. The validity of the proposed approach will have to be confirmed in future clinical studies. Nonetheless, the results herein reported suggest that our approach could be utilized in clinical assessments where other techniques have failed to capture the impact of soft tissue damage causing back pain.