Participants and ethical approval

Eighteen participants, 13 male and 5 female (33.3 ± 9.2 years), participated in the study. The inclusion criteria for participation was the absence of neurologic, cardiovascular or metabolic disorders. Furthermore, they had no previous surgeries, current or recurring pain, or injury on their dominant side. The study protocol was approved by the U.S. Army Research Laboratory Institutional Review Board and all participants gave their written informed consent in accordance with the Helsinki Declaration.

Experimental protocol

Participants performed two independent and discrete movements as a part of the research study: knee extension and elbow flexion of the dominant limb. The dominant limb was ascertained by asking the participant which limb they would use to kick or throw a ball. In the present study, all participants were right-side dominant. For the knee extension movement, participants were seated in a stationary chair with their foot at least 10 cm from the ground and a mass of 2.3 kg was attached to their ankle. A surface EMG electrode (B&L Engineering, Tustin, CA; Ag/AgCl, circular 10mm diameter, interelectrode distance 35mm, Gain 330x, >100MΩ, CMMR 95db, 10 Hz-3.13 kHz bandwidth) was attached to the vastus lateralis approximately 2/3 the distance between the anterior superior iliac spine and the lateral side of the patella. Manual muscle tests for simple hip flexion and leg abduction were performed to limit cross-talk. For the elbow flexion movement, participants were seated in a chair with their elbow supported in 90 degrees of flexion and a 2.3 kg mass attached to their wrist. A surface EMG electrode was attached midway between the lateral epicondyle and the acromion on the the short head of the biceps brachii as determined via palpation. Prior to the placement of the EMG electrode for both movements, the skin was lightly abraded and cleaned with a 70% isopropyl alcohol solution. A 5 mm diameter adhesive pre-gelled Ag/AgCl surface electrode was placed on the subject’s ipsilateral patella as a ground. All EMG was pre-amplified with a 330x gain and A/D converted at 2048 Hz. After placement of the EMG electrodes, the process for both movements was similar. The participant was asked to perform three repetitions of the specific movement at a self-selected pace separated by at least 60 seconds of rest. No attempt was made to standardize movement speed or EMG signal-to-noise ratio as the goal of the present analysis is to determine an algorithm or class of algorithms which are valid for use in a variety of conditions. Five trials were lost due to equipment malfunction, rendering 103 trials of experimental EMG data for analysis. Since this same data was used for the creation of the simulated EMG, this resulted in a net total of 206 time series’ of data for analysis.

Visual electromyography onset detection

A custom computer program was written in the R programming language which enabled all three researchers to visually determine and record EMG onset in a randomized and double-blind fashion. Within each researcher, the visual determinations were separated by at least 24 hours and the inter-determination period was never longer than 7 days. Prior to researcher evaluation of the EMG, the signal was bandpass filtered (10–1,000 Hz) in line with previous research to remove any substantial signal artifact [13]. The reviewing researchers were instructed to evaluate the time series for what they perceived to be the onset of muscle activity. Each researcher evaluated every trial twice and was blinded to both the study identifier as well as the movement performed. The reliability between and within investigators was high (ICC: 0.88 and 0.92, respectively); therefore, the mean of the six visual detections (two from each researcher) was used as the ‘gold standard’ from which to evaluate the algorithms [5, 13].

Creation of simulated EMG

The experimental data collected was altered to obtain a known onset of EMG signal. A half-second increment of data (1024 samples) was extracted from the obtained signal in a trial prior to muscle contraction. During this period the muscle was always quiescent. Since the greatest singular investigator visual onset deviation from visual ‘gold standard’ was 0.32 seconds, active EMG was obtained 0.5 seconds after the visual ‘gold standard’ determination. The length of active EMG was 1 second (2048 samples). Thus, the length of all simulated trials was 1.5 seconds (3072 samples) with an objective EMG onset at 0.5 seconds. While the onset in simulated EMG is objective, this method renders a data series with an artificially profound onset since it does not contain the gradual orderly recruitment of different motor units. However, the inability to sufficiently determine an onset in the simulated EMG indicates that an algorithm is inadequate even when a clear EMG signal change occurs.

Algorithmic electromyography onset detection

A net total of 605 algorithm iterations were tested (Table 1 and Table 2). Prior to analysis by any algorithm, the raw EMG waveform was bandpass filtered (10–1,000 Hz) to remove signal artifact [13]. Three classes of standard algorithms were examined (Table 1), including sixty-four iterations of the commonly applied linear envelope methodology [4]. The linear envelope was also tested with the Teager-Kaiser Energy Operator (TKEO) pre-processing step since this has been shown to increase accuracy [5, 6]. The Sample Entropy (SampEn) algorithm has also been shown to have accuracy similar to TKEO while being more robust to artifact [7]. In addition to testing these 129 standard EMG onset algorithms, 476 statistical algorithms for time series analysis were applied for use in EMG onset detection (Table 2).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Standard EMG onset detection methodologies examined and the iterative settings used for each methodology.

https://doi.org/10.1371/journal.pone.0177312.t001

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Statistical EMG onset detection methodologies examined and the iterative settings used for each methodology.

https://doi.org/10.1371/journal.pone.0177312.t002

The first general set of algorithms arose from the changepoints package in R [20]. These algorithms test the time series for At Most One Change (AMOC) by examining either a change in the mean of the time series, the variance of the time series, or both. For all three of these sub-methods, the distribution of the data was assumed to be normal. The second set of statistical algorithms were the sequential and batch processing of parametric and non-parametric methods from the cpm package in R [21]. In batch processing, the data is always retrospective and changepoints are calculated from the data as a whole (i.e. all observations in the time series). In sequential processing, the individual data points are received and processed over time until a changepoint is detected (i.e. at each ordered observation, a decision is made whether a change has occurred). The models used in both of these methodologies assume statistical independence between data points in the series, a criteria which may not be strictly met with surface EMG, but was assumed to be true for the current purposes of EMG onset detection. The individual methodologies applied from both batch and sequential processing can be viewed in Table 2. The third set of statistical algorithms were a Bayesian analysis of change points in a time series [22, 23] as implemented in the bcp package in R [23, 24]. As opposed to other methods tested in this study, Bayesian procedures do not produce a point estimate of the EMG onset; instead, a probability distribution is produced and the researcher can select their probability level for onset determination. In this implementation, the posterior means are updated after each partition within the time series. In the present study, the parameter of the prior (hyperparameter γ or p₀) is systematically varied so that the ability to detect a small number of changes (p₀ low) and a large number of changes (p₀ high) are examined [22]. In practice, the parameter can alter the sensitivity and specificity of the algorithm for EMG onset detection. The method also has the capability of accounting for differing levels of signal-to-noise (hyperparameter w₀) which can theoretically be altered based upon the characteristics of the underlying signal. This hyperparameter of the algorithm was set at 0.2 as prescribed by previous work in Bayesian changepoint analysis [22, 25].

Each general algorithm type has a number of parameters which alter how EMG onset is determined and defined. The range of parameters within each algorithm are adapted from the methods typically employed in the EMG literature (in the case of standard algorithms) or are a systematic exploration of the iterative changes available within individual statistical packages (in the case of statistically-oriented algorithms).

Statistical analysis & algorithm evaluation

Given the large number of EMG onset algorithms examined, an iterative process was used to down-select appropriate algorithms for further analysis. The down-select process was identical for experimental and simulated EMG except for step 3, which was omitted for simulated EMG. The followings steps were used to down-select algorithms: 1) the root mean square error (RMSE) between algorithm-determined EMG onset and visual onset was calculated and algorithms with the highest mean RMSE (top 90% of the 605 algorithm iterations) were removed; 2) algorithms which either detected no EMG onset or detected EMG onset at the first index point more than 25% of the time were removed; 3) for experimental EMG, a linear regression model with algorithm-detected onset and EMG signal-to-noise ratio were used to predict visual onset detection and any algorithms which had statistically significant (p < 0.05) effects for signal-to-noise ratio were removed from further analysis. The signal-to-noise ratio was calculated as the amplitude of the signal during ‘quiet’ EMG in the first 500 ms of data collection in ratio to the maximum amplitude observed during the trial. Despite the uncomplicated nature of the movements performed, the low resistance (2.3 kg) resulted in a reasonable range of signal-to-noise ratios (1.8–78.4). Step 3 was performed with the experimental EMG as the goal of the present investigation was to determine a set of algorithms which produce accurate results independent of signal quality. Any attempt to replicate this regression procedure for the simulated EMG would have resulted in an invalid analysis due to a lack of variance in the dependent variable (e.g. the variance of objective onset in simulated EMG is essentially ‘0’ since all EMG turns ‘on’ at 0.5 seconds).

The remaining algorithms for both experimental and simulated EMG datasets were assessed by determining the mean difference between the algorithm’s determined onset and the ‘gold standard’ visual onset (for experimental data) or objective onset (for simulated data) and the associated parametric 95% confidence intervals. In each case, the accuracy of the algorithm was assessed as the mean difference’s proximity to ‘0’ and the reliability was assessed by examining the width of the 95% confidence intervals (more narrow intervals indicate greater algorithm reliability). All statistical analyses were performed in R [26].

Bayesian changepoint analysis

The Bayesian changepoint analysis implemented in the current study is based on the original work by Barry and Hartigan [22] and extended as the R package bcp [23]. These analytical methods have been applied to a wide variety of fields, from genomics [24] to climate change [27] and investigations of the National Hockey League demographics [28]. In general, the analysis provides a posterior probability for the presence of a changepoint within a given partition of the time series. The posterior probability is updated after each partition is iterated and are conditional on the current partition. Two hyperparameters within Bayesian changepoint analysis can be ‘tuned’ to increase their efficacy for a given situation: p₀ and w₀. The exact derivation and use of these hyperparameters is detailed in the original work by Barry and Hartigan [22]; in the context of EMG onset, higher p₀ values are effective when there are many changepoints to detect and higher w₀ values are effective when the signal-to-noise ratio of the underlying EMG signal is low [22, 23]. Both p₀ and w₀ are bounded from 0–1.

In the confines of the present study, w₀ was not systematically altered and the default 0.2 level was used as prescribed by previous work [22, 25]. The p₀ hyperparameter was iteratively changed from 0–1 (Table 2). Given the current study which only examined a singular “EMG onset” or changepoint in a time series, it should be unsurprising that p₀ = 0.0 rendered the most reliable and accurate results. The present study also iteratively examined the most appropriate cut-point for the posterior probability, with various probabilities being more reliable in different data sets (experimental vs. simulated, see Figs 3 and 4). In practice, the selection of an appropriate posterior probability set-point may also be guided by the underlying theory and belief of the researcher regarding the desired level of confidence in the data.

The clearly superior performance of Bayesian changepoint analysis for the determination of surface EMG onset in both experimental and simulated data indicates that this methodology deserves greater study. Research which examines multiple muscle onsets and offsets in dynamic tasks should provide evidence that the p₀ hyperparameter needs to be varied depending on the type of EMG data collected. The w₀ hyperparameter may be able to be altered depending upon signal quality or the signal-to-noise ratio of the underlying EMG data. The differential fixation of these two hyperparameters will also impact the overall algorithm accuracy and reliability. In theory, a function could be derived which modifies the hyperparameters based upon the underlying characteristics of the EMG signal being evaluated. Future research should explore this possibility as it may allow for the most reliable and flexible analysis of EMG onset across laboratories and in clinical applications.

Future work

Many of the novel algorithms tested in the current manuscript arise from areas of econometrics, genomics and statistical changepoint analysis. The superior performance of the Bayesian changepoint algorithms reinforce the notion that biomechanics and motor control researchers should continually assess the state-of-the art in adjacent or seemingly unrelated fields to leverage the computational or experimental advances in those fields. Computational work in the areas of non-linear dynamics [29, 30], recursive neural networks [31, 32] and network dynamics [33] may be especially fruitful for extracting EMG onset timing.