COMPUTER DETECTION APPROACHES FOR IDENTIFICATION OF PHASIC ELECTROMYOGRAPHIC (EMG) ACTIVITY DURING HUMAN SLEEP

Abstract

BACKGROUND

Examination of spontaneously occurring phasic muscle activity from the human polysomnogram may have considerable clinical importance for patient care, yet most attempts to quantify the detection of such activity have relied upon laborious and intensive visual analyses. We describe in this study innovative signal processing approaches to this issue.

METHODS

We examined multiple features of surface electromyographic signals based on 16,200 individual 1-second intervals of low impedance sleep recordings. We validated which of those features most closely mirrored the careful judgments of trained human observers in making discriminations of the presence of short-lived (100-500 msec) phasic activity, and also examined which features provided maximal differences across 1-second intervals and which features were least susceptible to residual levels of amplifier noise.

RESULTS

Our data suggested particularly promising and novel features (e.g., Non-linear energy, 95^th percentile of Spectral Edge Frequency) for developing automated systems for quantifying muscle activity during human sleep.

CONCLUSIONS

The EMG signals recorded from surface electrodes during sleep can be processed with techniques that reflect the visually based analyses of the human scorer but also offer potential for discerning far more subtle effects, Future studies will explore both the clinical utility of these techniques and their relative susceptibility to and/or independence from signal artifacts.

1.1 INTRODUCTION

Polysomnography (PSG), the recording of multiple physiologic signals during sleep, generates huge amounts (typically 500 megabytes or more for an overnight sleep period) of electrophysiologic data that are typically acquired digitally but remain analyzed visually by human expert scorers [1]. Of PSG parameters typically recorded, the electroencephalogram (EEG), by far, has been the most extensively examined using advanced signal processing techniques for data reduction, including approaches such as spectral analysis using the Fast Fourier Transform (FFT) [2], period amplitude analysis [3], and non-linear dynamics employing estimates of dimensional complexity [4]. Although such approaches have stimulated many exciting studies of the functional significance of the sleep EEG [5, 6], other components of the human PSG has received far less attention from the signal processing perspective. Some work, for example, has attempted to automate identification of the eye movements of rapid eye movement sleep (REM) [7], whereas other studies have attempted to quantify muscle activity from surface electromyography (EMG) during sleep [8-11]. In recent years, greater attention has focused on the analysis of EMG signals derived from sleep, because of known aberrations of such activity during REM as a component of some neurodegenerative conditions [12-16]. We describe here the initial derivation of an unsupervised, computer-based approach for quantifying surface EMG during sleep that is both compatible with discriminations provided by the trained human scorer, as well as potentially extending such quantification to incorporate unique signal processing features.

2.0 METHODS

2.1. Overall Approach

To establish a viable digital approach of the human PSG for the quantification of EMG activity, we initially relied upon judgments of trained visual scorers. Five different individuals (A,B,C,D,E), each representing different levels of experience, independently evaluated surface EMG signals from a human PSG recording, as recorded with a low impedance (< 10,000 ohms), bipolar derivation placed directly over a limb muscle group frequently associated with discharge in sleep (anterior tibialis). Our general strategy was to use a consensus of human scorers’ determinations of the presence or absence of phasic activity for each 1-second interval. Although EMG signals during sleep are frequently recorded from surface sensors located above other muscle groups as well (e.g., mentalis), such locations can be problematic for introducing certain kinds of artifacts (e.g., snoring) and would represent a more complex issue for human scorer discrimination, at least at this early stage of developing computer-based learning approaches. Scorers varied in experience level from an individual with over 30 years of experience in examining PSG to a lab assistant with very modest (< 6 months) experience. Our rationale for inclusion of such a wide array of scorer expertise was to enhance generalizeability of any algorithm that could result from our analyses.

Sleep (PSG) data were collected on an Embla (MedCare, Bloomfield, CO) Model N7000 data acquisition unit using the sleep data collection software program Somnologica ® 2.0. PSG data were converted from Somnologica .edf format, for EMG feature processing, via the computer software program MATLAB (version 7.8 R2009a) using the biosig toolbox version 2.54 (Schloegl A-Graz University of Technology, Graz, Austria) and were digitized at a sampling rate of 200 Hz.

Visual analyses occurred on 17″ monitors. Each of the 5 scorers was asked to evaluate a total of 16,200 seconds (4.5 hours) of bilateral EMG signals distributed approximately equally from REM and non-REM sleep, marking each second (via drag-and-click function) for the presence or absence of a phasic muscle discharge of 4 times the surrounding background activity and lasting between 100 and 500 msec. This definition resembles what we have used previously for visual analyses of the phasic EMG during sleep [17]. Screen display was held constant at 10 seconds per viewing window. Because we relied on a 1-second epoch to define presence or absence of phasic activity, scorers were instructed to score an event as present or absent for the 1-second interval, rather than attempt to quantify the number of phasic events within the given interval (see Figure 1), thus allowing for repetitive returns to baseline during that interval. Each scorer was provided an unmarked recording and had no access to the judgments made by any of the other visual scorers during the entire process.

Figure 1.

Somnologica ® 2.0 RemLogic plot of a 5-Second segment of Right Leg EMG data with phasic activity labeling from Scorer B. Phasic activity is demarcated with shaded dotted boxes and “P_RIGHT” labels.

2.2. Computer-based Feature Extraction from EMG Recordings

Feature extraction was used to translate EMG signal information into a compact quantitative format and characterize signal information for the identification of seconds with and without phasic muscle events. Exhaustive review of prior EMG sleep signal analysis, bio-signal feature extraction techniques, and statistical analysis methods were used to compile a library of 15 features, listed below. Formulae associated with the calculations are numbered and shown in bolded parentheses. The final data matrix was obtained by evaluating feature values for each data channel in one second epochs using eight equal non-overlapping moving windows.

1. Relative EMG Frequency Power (Prel): a frequency domain feature that provides a sub-band analysis of the high frequency EMG signal components (frequency band [12.5;32 Hz ]) [18] (sampled at 200Hz) with the power spectra (P(f)) extracted using the Fast Fourier Transform (FFT) [19, 20]

$$\mathit{Prel}=\frac{P\left(\right[12.5-32\mathit{Hz}\left]\right)}{P\left(\right[8-32\mathit{Hz}\left]\right)}.$$ (1)

2. Spectral Edge Frequency 95^th percentile (SEF 95): the frequency up to which 95 percent of the total signal power is accumulated [21].

$$\underset{0}{\overset{\mathit{SEF}95}{\int }}P\left(f\right)\mathit{df}=0.95\underset{0}{\overset{f_{s∕2}}{\int }}P\left(f\right)\mathit{df}$$ (2)

where f_s is the sampling frequency

3. Skewness (Skew): a time domain feature that measures the asymmetry of the probability distribution of the EMG signal amplitude [22].

$$\mathrm{Skew}=\frac{\frac{1}{M}\sum _{i=1}^M{\left(x\right(i)-\stackrel{‒}{x})}^3}{{\left(\sqrt{\frac{1}{M}\sum _{i=1}^M{\left(x\right(i)-\stackrel{‒}{x})}^2}\right)}^3},$$ (3)

with M representing the number of data samples of the EMG signal in an 1-second epoch interval and $\stackrel{‒}{x}$ symbolizing the sample mean $\stackrel{‒}{x}=\frac{1}{M}\sum _{i=1}^{M}x\left(i\right)$ within that interval,

4. Variance (S²): the distance, spread, between a data sample with respect to the mean [22].

$${\mathrm{S}}^2=\frac{\sum _{i=1}^M{\left(x\right(i)-\stackrel{‒}{x})}^2}{M-1},$$ (4)

5. Kurtosis (Kurt): a measure of the peakedness or flatness of the probability distribution of the signal amplitude [23].

$$\mathrm{Kurt}=\frac{\frac{1}{M}\sum _{i=1}^M{\left(x\right(i)-\stackrel{‒}{x})}^4}{{\left(\frac{1}{M}\sum _{i=1}^M{\left(x\right(i)-\stackrel{‒}{x})}^2\right)}^2},$$ (5)

6. Entropy (Ent): an information domain feature that calculates the amount of uncertainty or unpredictability of the EMG signal amplitude.

$$\mathrm{Ent}=-\sum _{j=1}^n\frac{{\mathit{bin}}_j}{M}\log \left(\frac{{\mathit{bin}}_j}{M}\right),$$ (6)

with M symbolizing the length of the data signal, n representing the number of bins, the optimal bins being obtained from the Freedman-Diaconis rule [24], to estimate the histogram of the data signal with bin_j indicating the number of data samples from x(i) in the j^th histogram bin [25].

7. Mobility (Mobi): a time domain feature that measures the relative average slope of the EMG signal. It is expressed as the standard deviation (std) of the slope (signal’s first derivative dx/dt) with reference to the standard deviation of the signal amplitude [25].

$$\mathrm{Mobi}=\frac{\mathit{std}\left(\frac{d\left(x\right)}{\mathit{dt}}\right)}{\mathit{std}\left(x\right)},$$ (7)

where for the discrete EMG signal x(i), $\mathit{std}\left(x\right)=\sqrt{\frac{\sum _{i=1}^M{\left(x\right(i)-\stackrel{‒}{x})}^2}{M-1}}$ and $\mathit{std}\left(\frac{d\left(x\right)}{\mathit{dt}}\right)=\sqrt{\frac{\sum _{i=1}^{M-1}{\left(f_{\mathit{sampl}}\left(x\right(i+1)-x(i\left)\right)-\frac{1}{M-1}\sum _{i=1}^{M-1}{f}_{\mathit{sampl}}\left(x\right(i+1)-x(i\left)\right)\right)}^2}{M-2}}$, using a first order approximation of the derivative.

8. 75^th Amplitude Percentile (75_Amp): the amplitude value below which 75% of the total EMG signal amplitude resides [25]. So, the value separates lowest 75% and highest 25% of the data. It is also called upper quartile or third quartile.

$$\mathit{card}\left\{x\right(i)\mid x(i)<75\_\mathit{Amp}\}=\frac{75.M}{100},$$ (8)

where M is the number of samples x(i) of the EMG signal in one epoch and card represents the number of elements within the sample set (set’s cardinality).

9. Complexity (Comp): the ratio of the mobility (Mobi) of the first derivate of the signal to the mobility of the signal amplitude. Complexity expresses the average EMG wave-shape in relation to a pure sine wave [25].

$$\mathrm{Comp}=\frac{\mathit{Mobi}(\mathit{dx}∕\mathit{dt})}{\mathit{Mobi}\left(x\right)},$$ (9)

10. Mean Absolute Amplitude (MAA): a time domain feature that measures the absolute value of the mean EMG amplitude [26].

$$\mathrm{MAA}=\frac{1}{M}\sum _{i=1}^M\mid x\left(i\right)\mid ,$$ (10)

11. Curve Length (L): the sum of the value of the first order differences of the EMG signal amplitude values [27].

$$\mathrm{L}=\sum _{i=1}^{M-1}\mid x(i+1)-x\left(i\right)\mid ,$$ (11)

12. Mean Energy (MnE): a time domain feature that measures the squared EMG signal amplitude [27].

$$\mathrm{MnE}=\frac{1}{M}\sum _{i=1}^{M}x{\left(i\right)}^2,$$ (12)

13. Zero Crossings (ZC): defined as the number of crossings of the EMG signal over the ordinate, where the axis equals zero [27].

$$\begin{array}{cc}\hfill {\mathit{ZC}}^{\prime }& \left(i\right)=[\begin{array}{c}1,x\left(i\right)\le 0\cap x(i+1)>0\hfill \\ 1,x\left(i\right)\ge 0\cap x(i+1)<0\hfill \\ 0,\mathit{otherwise}\hfill \end{array}\phantom{]}\hfill \\ \hfill & \mathit{ZC}=\sum _{i=1}^{M-1}{\mathit{ZC}}^{\prime }\left(i\right),\hfill \end{array}$$ (13)

14. Average Nonlinear Energy (NE): a non-linear feature that is sensitive to signal fluctuations in the time and frequency domain, with respect to the following non-linear operator (NLO):

$$\mathit{NLO}\left[i\right]=x{\left(i\right)}^2-x(i-1)x(i+1)$$

the NLO is weighted with a Hanning window and then the mean/Average Nonlinear Energy (NE) is calculated as follows:

$$\mathit{NE}=\frac{1}{N}\sum _{i=1}^N{\mathit{NLO}}_w\left[i\right],$$ (14)

where, NLO_w is the Hamming windowed version of the non-liner operator, NLO, with N being the data epoch [27].

15. Spectral Entropy (SE): defined as the amount of uncertainty or unpredictability of the EMG signal in the frequency domain [27].

$$\mathit{SE}=-\Sigma P\left(f\right){\log }_{2}P\left(f\right),$$ (15)

2.3 Performance Models Comparing Features

We evaluated the relative performance of each feature using three analytic approaches. First, we examined the relative accuracy afforded by each feature by examining percentage agreement for each feature by use of an unsupervised, probabilistic model of likelihood of class assignment in relation to human scorer consensus. Accuracy was defined as the sum of the percentage of correctly (i.e., consensually) identified seconds with phasic activity (true positive) plus the percentage of correctly (i.e., consensually) identified seconds without phasic activity (true negative). Consensus was defined as unanimous agreement among all 5 human scorers. These calculations were made separately for left and right leg channels to provide internal replication of the performance of various features.

Unsupervised classification of 1-second epochs with phasic versus non-phasic activity were conducted using Gaussian Mixture Modeling (GMM) [28]. GMMs are a member of the probabilistic model based clustering framework. GMMs assume that the data come from a multivariate finite mixture model of the form

$$p\left(\mathrm{x}\right)=\sum _{k-1}^K{\pi }_k{N}_k(\mathrm{x};{\mu }_k,{\Sigma }_k),\sum _{k=1}^K{\pi }_k=1,\text{and}{\pi }_k\ge 0\text{for}k=1,2,\dots ,K$$

where K is the number of components , π_k are the mixing proportions (weights) and N_k (x;μ_k, ∑_k) are multivariate normal distributions of the form

$$N_k(\mathrm{x};{\mu }_k,{\Sigma }_k)=\frac{1}{{\left(2\pi \right)}^{\frac{n}{2}}{\mid {\Sigma }_k\mid }^{\frac{1}{2}}}\exp \left\{-\frac{1}{2}{(\mathrm{x}-{\mu }_k)}^T{\Sigma }_k^{-1}{(\mathrm{x}-{\mu }_k)}^T\right\}$$

where n is the dimension of the feature vector x , μ_k is the mean vector and ∑_k is the covariance matrix.

Estimates for the model parameters were obtained using the expectation maximization (EM) algorithm, a common approach in GMM parameter selection, which is an iterative algorithm used to determine the maximum likelihood (other criteria can be used but the maximum likelihood is the most common) of the data given the parametric model [29]. Once the parameters are determined each object is assigned to one the K clusters (components representing phasic or non-phasic activity) on the basis of estimated posterior probabilities of cluster membership. In other words the object x , feature value of each 1-second epoch, was assigned to cluster i if

$${\pi }_i{N}_i(\mathrm{x};{\mu }_i,{\Sigma }_i)\ge {\pi }_j{N}_j(\mathrm{x};{\mu }_j,{\Sigma }_j)\text{for all}i\ne j;j=1,2,\dots ,K$$

A second approach to evaluating the relative utility among features was to compare parametric values across consensually agreed upon seconds containing phasic EMG activity. Because visual analyses were limited to a simple binary classification, we reasoned that the additional information incorporated by some features would result in higher coefficients of variation (COVs) $(\frac{\mathit{std}}{\stackrel{‒}{x}}\times 100)$ for those features relative to other features. Higher values would be suggestive of potentially greater sensitivity to fluctuations in occurrence (either as density within the time domain and/or amplitude within the voltage domain) of muscle activity that would be difficult to detect visually. For example, Figure 1 demonstrates three distinct seconds as identified as demonstrating phasic activity, but the visual appearance of these recording segments varied widely. Ideally, any computer-derived feature would be sensitive to such variation across discrete 1-second intervals and demonstrate relatively larger, rather than smaller, COVs.

A final approach for comparison was to determine the relative success of each feature to deal with the low amplitude, high frequency, residual noise present within the N7000 recording system. Even within the context of low impedance recordings and the high performance amplification with a substantial common mode rejection ratio (CMRR) > 90dB at 60 Hz for the Embla N7000 system, some non-physiologic, low voltage (< 1 uV) signals are always detectable. The CMRR measures the tendency of the amplifier to reject input signals that are common to both EMG electrodes and a high CMRR (>90dB) is assumed to suppress signals common to both electrodes sites, such as from power sources, electromagnetic devices and muscles distant from the electrode radius [30]. Assuming the latter, we reasoned that, for seconds with consensus agreement on absence of phasic activity, the ratio of such so-called “residual noise” across left and right leg EMG channels would approach unity. Features that showed ratios considerably higher or lower than 1.0 would be considered to have greater susceptibility to spurious signals, regardless of their origin.

3.0 RESULTS

3.1 Visual Analyses

Approximately 10% of all seconds evaluated contained phasic muscle activity. The percentages of 1-second intervals detected for each of the 5 scorers were: 7.67, 7.90, 8.80, 9.44 and 12.70. The overall consensus (i.e., the proportion of 1-second intervals in which all 5 scorers agreed on the presence or absence of phasic muscle activity was very high (93.52%; Kappa = .818, p < .0001). Because we were able to establish high reliability among scorers with regards to the presence and absence of phasic activity, we limited further our machine-learning analysis to only those 1-second intervals with consensus agreement on presence (N = 1,090) or absence (N = 14,060) of EMG phasic activity.

3.2 Relative Classification Performance of Computer-derived Features from Unsupervised Learning Models

Tables 1 and 2 present results evaluating the performance of each of the 15 features for the unsupervised models derived from the EMG signals using the GMM algorithm. With few exceptions, most showed substantial differentiation between seconds with and without consensually agreed phasic muscle activity. Nine of the features demonstrated accuracy above 90% and of these, only Comp (complexity) showed substantial discrepancy between seconds identified accurately with phasic activity (true positives) versus seconds identified accurately as not containing phasic activity (true negatives).

Table 1.

Accuracy (% agreement) for Consensus Visual Scoring (left leg)

Feature Name (Number)a	Accuracyb	TPc	TNd
Relative High Frequency Power (1)	63.530	13.849	66.917
Spectral Edge Frequency 95th Percentile (2)	91.838	95.723	91.573
Skewness (3)	89.862	68.839	91.295
Variance (4)	90.265	97.352	89.782
Kurtosis (5)	88.316	83.707	88.629
Entropy (6)	45.646	50.917	45.287
Mobility (7)	89.537	45.214	92.559
Amplitude 75th Percentile (8)	96.829	91.853	97.168
Complexity (9)	91.786	20.774	96.626
Mean Absolute Amplitude (10)	92.228	96.945	91.906
Curve Length (11)	92.488	97.759	92.128
Energy (12)	90.265	97.352	89.782
Zero Crossing (13)	73.447	56.823	74.580
Non-Linear Energy (14)	92.995	97.556	92.684
Spectral Entropy (15)	94.957	86.354	95.544

^aRelevant formulae are shown in parentheses (see text)

^bAccuracy = % of correct phasic and non-phasic GMM predictions compared to consensus visual scoring

^cTP = true positives (% of seconds with phasic activity assessed visually that were correctly classified by the GMM for that feature)

^dTN = true negatives (% of seconds without phasic activity assessed visually that were correctly classified by the GMM for that feature).

Table 2.

Accuracy (% agreement) for Consensus Visual Scoring (right leg)

Feature Name (Number)a	Accuracyb	TPc	TNd
Relative High Frequency Power (1)	64.713	17.028	68.879
Spectral Edge Frequency 95th Percentile (2)	90.410	97.496	89.791
Skewness (3)	89.820	66.611	91.848
Variance (4)	88.747	98.164	87.925
Kurtosis (5)	86.347	84.641	86.496
Entropy (6)	59.134	96.350	13.419
Mobility (7)	89.311	33.222	94.210
Amplitude 75th Percentile (8)	95.789	94.825	95.873
Complexity (9)	90.585	20.868	96.675
Mean Absolute Amplitude (10)	90.370	97.997	89.704
Curve Length (11)	90.813	98.498	90.141
Energy (12)	88.774	98.164	87.954
Zero Crossing (13)	71.473	54.090	72.991
Non-Linear Energy (14)	92.073	97.997	91.556
Spectral Entropy (15)	97.425	84.641	98.542

^aRelevant formulae are shown in parentheses (see text)

^bAccuracy = % of correct phasic and non-phasic GMM predictions compared to consensus visual scoring

^cTP = true positives (% of seconds with phasic activity assessed visually that were correctly classified by the GMM for that feature)

^dTN = true negatives (% of seconds without phasic activity assessed visually that were correctly classified by the GMM for that feature).

Visualization of the agreement/overlap between consensus scoring and GMM classifications were displayed using histogram plotting for three selected features (Figures 2-7; right leg data shown; results were similar using left leg). This information follows directly from Table 2 and illustrates the accuracy of the unsupervised GMM algorithm in classifying absence of phasic and phasic EMG activity with respect to consensus scoring. Histograms for a poorly performing feature, Relative High Frequency Power (formula 1), are displayed in Figures 2 and 3, with overlap values of 68.9% and 17.0% for absence of phasic and presence of phasic EMG activity, respectively. Results for a mid-level performing feature, Zero-Crossing (formula 13) are displayed in Figures 4 and 5 with overlap values of 73.0% and 54.1% for absence of phasic and presence of phasic EMG activity, respectively. Lastly, histograms for an optimally performing feature, Variance (formula 4), are displayed in Figures 6 and 7 with overlap values of 87.9% and 98.2% for absence of phasic and presence of phasic EMG activity, respectively.

Figure 2.

Histogram plots of the absence of phasic activity for consensus visual scoring and the GMM algorithm for a poorly performing feature: Relative High Frequency Power (1) (Right Leg). Purple color indicates overlapping sections of the consensus scoring and GMM classification histograms.

Figure 7.

Histogram plots of the presence of phasic activity from consensus visual scoring and GMM for an optimally performing feature: Variance (4) (Right Leg). Purple color indicates overlapping sections of the consensus scoring and GMM classification histograms.

Figure 3.

Histogram plots of the presence of phasic activity agreement labeling from consensus visual scoring and the GMM algorithm for a poorly performing feature: Relative High Frequency Power (1) (Right Leg). Purple color indicates overlapping sections of the consensus scoring and GMM classification histograms.

Figure 4.

Histogram plots of the absence of phasic activity for consensus visual scoring and the GMM algorithm for a mid-level performing feature: Zero Crossing (13) (Right Leg). Purple color indicates overlapping sections of the consensus scoring and GMM classification histograms.

Figure 5.

Histogram plots of the presence of phasic activity agreement labeling of the consensus visual scoring and GMM algorithm for a mid-level performing feature: Zero Crossing (13) (Right Leg). Purple color indicates overlapping sections of the consensus scoring and GMM classification histograms.

Figure 6.

Histogram plots of the absence of phasic activity for consensus visual scoring and the GMM algorithm for an optimally performing feature: Variance (4) (Right Leg). Purple color indicates overlapping sections of the consensus scoring and GMM classification histograms.

3.3 Coefficients of Variation (COVs) across Features

Table 3 presents COVs established for left and right leg separately for various EMG features for consensually agreed upon seconds with phasic activity. Although some differences occurred for particular features in left and right leg recordings, features (2) (SEF, Spectral Edge), (4) (S², Variance), (12) (MnE, Energy), (14) (NE, non-linear energy) and (15) (SE, Spectral Entropy) showed consistently higher COVs.

Table 3.

Coefficients of Variation (COV) across 1-second Intervals for Consensually Agreed Seconds with Phasic Activity

Feature Name (Number)a	COV (Left)	COV (Right)
Relative High Frequency Power (1)	77.3	80.5
Spectral Edge Frequency 95th Percentile (2)	711.5	916.0
Skewness (3)	75.5	74.1
Variance (4)	880.2	807.9
Kurtosis (5)	129.1	118.6
Entropy (6)	58.8	45.0
Mobility (7)	60.8	63.7
Amplitude 75th Percentile (8)	292.7	329.3
Complexity (9)	204.1	222.1
Mean Absolute Amplitude (10)	188.9	183.8
Curve Length (11)	200.1	200.1
Energy (12)	912.1	809.3
Zero Crossing (13)	51.1	49.1
Non-Linear Energy (14)	613.5	1308.4
Spectral Entropy (15)	1602.3	1343.5

^aRelevant formulae are shown in parentheses (see text)

3.4 “Residual Noise” across Features

Table 4 presents the ratio between left and right consensually agreed upon seconds without phasic activity. Ratios approximating 1.0 suggested features that were relatively robust to residual noise within the recording system. Selected features (1, 3, 5, 6, 7, 13) all approximated unity. Several features with promising discrimination in the unsupervised models (Tables 1 and 2) and/or in the COV analysis (Table 3), e.g., SEF, Spectral Edge (2); S², Variance (4); SE, Spectral Entropy (15), showed substantial divergence, suggesting a higher susceptibility to artifact.

Table 4.

Ratio of Left-to-Right Feature Activity for Consensually Agreed Seconds without Phasic Activity

Feature Name (Number)a	Ratio (L/R)
Relative High Frequency Power (1)	0.82
Spectral Edge Frequency 95th Percentile (2)	0.13
Skewness (3)	0.98
Variance (4)	0.12
Kurtosis (5)	1.04
Entropy (6)	0.95
Mobility (7)	0.97
Amplitude 75th Percentile (8)	0.68
Complexity (9)	0.81
Mean Absolute Amplitude (10)	0.62
Curve Length (11)	0.70
Energy (12)	0.12
Zero Crossing (13)	0.98
Non-Linear Energy (14)	0.46
Spectral Entropy (15)	0.20

^aRelevant formulae are shown in parentheses (see text)

4.0 DISCUSSION

When based solely on accuracy in relation to visually analyzed EMG recordings, a number of different computer-derived features (SEF, S², Kurt, 75_Amp, MAA, L, MnE, NLE, SE) appeared promising approaches to quantifying the human sleep EMG signal. However, when examined from standpoint of maximizing differences among consensually agreed upon seconds with phasic activity (Table 3), only 5 of these features (SEF, S², MnE, NLE, SE) retained highest variability, although these also appeared to be more susceptible to residual noise artifact (Table 4). Features promising lower influence of residual noise (Prel, Skew, Kurt, Ent, Mobi, ZC) showed reduced variability across consensually agreed upon seconds with phasic activity and/or lower accuracy in relation to consensually agreed upon seconds with and without phasic activity. These data suggest that further studies applying computer-derived metrics derived from digitally processed muscle activity during sleep should examine multiple features simultaneously to determine whether susceptibility to residual noise bears upon feature discrimination.

Recent interest in quantification of the EMG during sleep has been prompted, at least in part, by the recognition that a particular class of neurodegenerative disease (i.e., the synucleinopathies, including Parkinson’s disease, Dementia with Lewy Bodies, Multi-system Atrophy, idiopathic Rapid Eye Movement Sleep Behavior Disorder) are characterized by elevated muscle activity in sleep, particularly in REM, but often in Non-REM sleep as well [17, 31, 32]. Numerous investigators, including ourselves, have quantified muscle activity visually during sleep, typically following a variant of the manual (c.f., computer detected) approach outlined in the original study of Lapierre and Montplaisir [33]. Studies such as these have shown robust differentiation from controls [34, 35], progression of severity [36] or development of Parkinson’s disease [37, 38] over time, short-term stability across consecutive nights [39, 40], validation with video monitoring of motor behavior [41, 42] and spouse or caregiver questionnaires [43], and high synchronization of firing across muscle groups during REM [44]. Taken together, these diverse findings indicate that careful evaluation of muscle activity during sleep may hold important diagnostic and prognostic significance for this class of neurodegenerative diseases. However, given the tedious, time-consuming, and labor-intensive nature of such visual analysis, the foregoing findings also suggest the necessity of an automated approach to this process.

Our approach to quantification differs from those used by other investigators of the surface EMG in sleep in several important respects. First, although there is a long history of attempts to quantify muscle activity during sleep, our study is the initial attempt to extend automated approaches to include a complete library of techniques that incorporate several novel measures for EMG quantification, more typically applied for the sleep EEG but not for sleep-related muscle activity (e.g., non-linear dynamics, entropy). Early studies quantifying muscle activity of sleep employed cumbersome analog-to-digital hardware, which did little more than summate voltage repetitively to a pre-established criterion level [8]. Later work using 100-128 Hz digitization of EMG signals, focused largely on reducing muscle activity to a binary measure predicting desynchronization in the EEG [45] or examining activation levels over intervals of 20 seconds to examine associations between muscle tone and REM versus –Non-REM sleep [9]. The latter study’s reliance on variance measures anticipated to some extent the later variance-based approach of Burns et al. [10] who, using 3-second windows and a 200Hz sampling rate, defined high levels of mentalis activity as transient occurrence in the 5^th percentile of the variance distribution across all such intervals during the entire night of sleep. In another recent study, Ferri et al. [11] provided a detailed analysis of amplitudes of surface mentalis EMG signals within 1-second intervals and demonstrated differences in the voltage and time domain as a function of REM versus Non-REM sleep in healthy young controls, differences between younger controls and aged controls in REM sleep, and (in REM) differences between aged controls and patients with varied synucleinopathies. The approaches used by Burns et al. [10] and Ferri et al. [11] resembled several features examined here in our study in quantifying variance (formula 4) and amplitude (formulae 1, 8 and 10). By contrast, Mayer et al. [46] also employed an amplitude-derived smoothing function to determine a moving baseline, examined both voltage and time domains, but reported that slightly longer duration (0.5-10.0 sec; most less than 2.0 sec) elevations in muscle activity provided the clearest, clinically relevant differences.

A second way that our approach differs from at least some of the recent studies examining the sleep EMG is that investigators often have attempted to distinguish between short-duration (< 100 msec) activity representing motor unit activity and relatively longer duration (up to 5 or, in some cases, 30 seconds) tonic activity, presumably representing the recruitment of additional motor units within a muscle group [47]. Although the latter situation occurs, albeit variably, within the sleep EMG, the ability of human scorers to make this distinction visually (c.f., Figure 1B, pg 447 in [48] vs Figure 3, pg 2048 in [34]) can be difficult. Some automated digitization approaches [10, 11] simply allow the time domain to vary without arbitrary restrictions on duration of activity, whereas others make a binary differentiation of short versus long durations (e.g., [46]). Because phasic activity in a surface muscle recording most closely represents the fundamental activity of a firing motor unit, our validation has focused on phasic events (< 500 msec duration) occurring in the context of low impedance, high quality surface EMG recording free from artifacts and gross movement. Longer durations of motor activation [46], some resulting in behaviorally observable movement in which engagement of multiple motor units and involvement of multiple muscle groups are likely [47], might be expected to present a more complex challenge for analysis. Recordings with spurious signals, such as balistocardiographic artifact, also will represent an analytic challenge. Recordings from other muscle groups, such as the mentalis, which might be subject to different kinds of artifacts (e.g., inspiratory and/or expiratory vibratory noise), would represent yet another technical consideration. Our future work using the features analyses described in this study will pursue such issues.

Finally, our approach to automated signal processing of the human EMG signal subscribes to the concept that before any utilization of any system can be implemented at the level of the individual patient (i.e., by characterizing an entire overnight PSG), the most fundamental validation should mirror accurately the operation of a human visual scorer on a second-to-second, rather than a whole-night, basis. Ultimately, such an elemental step is essential before any attempt is made to apply complex signal processing to the vagaries clinical diagnosis of disease, an exceedingly complex issue in its own right. In the case of synucleinopathies, definitive disease definition of requires neuropathologic verification [49], although clinical criteria have been agreed upon consensually and are widely used [50]. Perhaps more relevant to the current effort is that for relatively new application of a potentially diagnostic tool like phasic sleep EMG, epidemiologic considerations involving sensitivity, specificity, positive predictive value and negative predictive value must all be examined on a population-wide basis. Clinical studies of the sleep EMG like those cited above vastly overselect for likelihood of synucleinopathic conditions and might be expected to be biased in favor of more accurate discrimination on a case-by-case (c.f., a second-by-second) basis. By way of example, although the prevalence of conditions like rapid eye movement sleep behavior disorder have been estimated at 1.6 cases per 100 normals in the general population [51] gross oversampling of such cases in clinical studies of the sleep EMG (e.g., 48 patients vs 25 controls [46]; 17 patients to 6 controls [43]; 80 patients to 80 controls [34]; 54 patients to 35 controls [35]) have the potential to create highly biased upwards estimates of diagnostic accuracy. It remains uncertain how well phasic EMG activity could discriminate incipient synucleinopathies in samples mirroring true prevalence of patterns of muscle activity during sleep observed in the general population. Initial validation of signal-processing methodology for phasic EMG should take place independently from such issues of case identification.

5.0 CONCLUSIONS

In summary, we have described in this study an approach for computerized detection of phasic EMG activity from surface electrodes during prolonged intervals of human PSGs. Several of the novel signal processing features tested here appeared to be promising for second-to-second validation with the trained human visual scorer, yet were not unduly influenced by system noise and retained the ability to detect a broad range of activity not otherwise detectable by the human scorer. Future studies will examine the relative resilience of such features to recording segments that will incorporate more diverse patterns of motor activation during human sleep.

HIGHLIGHTS.

• Surface recordings of phasic EMG in sleep may be prognostic for neurodegeneration

• Most attempts to quantify such activity have relied upon visual judgments

• We describe computer-derived, digitally based features for quantifying EMG in sleep

• Features reflect human scorers but discriminate beyond visually based approaches

• Several unique signal processing features are promising for future studies

Abbreviations

PSG

Polysomnography

EMG

Electromyography

EEG

Electroencephalography

REM

Rapid Eye Movement

COV

Coefficient of Variation

GMM

Gaussian Mixture Modeling

EM

Expectation Maximization

Prel

Relative High Frequency Power

SEF

Spectral Edge Frequency

S²

Variance

Kurt

Kurtosis

Ent

Entropy

Mobi

Mobility

75_Amp

75^th Amplitude Percentile

Comp

Complexity

MAA

Mean Absolute Amplitude

L

Curve Length

MnE

Mean Energy

ZC

Zero Crossings

NE

Nonlinear Energy

SE

Spectral Entropy