A Novel Method for Classification of Running Fatigue Using Change-Point Segmentation

lood lactate accumulation is a crucial fatigue indicator during sports training. Previous studies have predicted cycling fatigue using surface-electromyography (sEMG) to non-invasively estimate lactate concentration in blood. This study used sEMG to predict muscle fatigue while running and proposes a novel method for the automatic classi cation of running fatigue based on sEMG. Data were acquired from 12 runners during an incremental treadmill running-test using sEMG sensors placed on the vastus-lateralis, vastus-medialis, biceps-femoris, semitendinosus, and gastrocnemius muscles of the right and left legs. Blood lactate samples of each runner were collected every two minutes during the test. A change-point segmentation algorithm labeled each sample with a class of fatigue level as (1) aerobic, (2) anaerobic, or (3) recovery. Three separate random forest models were trained to classify fatigue using 36 frequency, 51 time-domain, and 36 time-event sEMG features. The models were optimized using a forward sequential feature elimination algorithm. Results showed that the random forest trained using distributive power frequency of the sEMG signal of the vastus-lateralis muscle alone could classify fatigue with high accuracy. Importantly for this feature, group-mean ranks were signi cantly di erent (p<0.01) between fatigue classes. Findings support using this model for monitoring fatigue levels during running. Keywords:surface-electromyography; blood lactate concentration; random forest; running; fatigue 1. Introduction Both recreational runners and running athletes can enhance their training outcomes by monitoring their fatigue level while training. The concept of muscular fatigue is usually described in two levels, or thresholds, where the rst threshold represents neuromuscular fatigue and the

support using this model for monitoring fatigue levels during running. Keywords:surface-electromyography; blood lactate concentration; random forest; running; fatigue 1. Introduction Both recreational runners and running athletes can enhance their training outcomes by monitoring their fatigue level while training. The concept of muscular fatigue is usually described in two levels, or thresholds, where the rst threshold represents neuromuscular fatigue and the second represents fatigue due to an imbalance between lactate production and lactate removal [1]. Most markers that are used to indicate fatigue during running, i.e., lactate or ventilatory parameters, need to be measured via blood samples or ventilation. However, due to the inconvenience of this during outdoor and recreational activities, neither of these are feasible to use in everyday training. Therefore, both recreational and competitive runners could be assisted by a new monitoring system that can predict fatigue level while running using noninvasive, portable instruments. Examples of such measurement units are electromyography devices, which in recent years have become small enough to be portable and possible to use inside clothing materials [2]. Previous studies have shown that features from an electromyography (EMG) signal can be used to detect fatigue thresholds both during cycling and running [3,4]. The features that have most frequently been used to calculate the fatigue thresholds are amplitude-based measures, such as the root Sensors2019,19, 4729; doi:10.3390 /s19214729 /journal/sensors

Sensors2019,19, 4729 2 of 17 mean square and integrated EMG, and frequency-based measures, such as mean power frequency or median frequency [3]. However, there are also studies that report a lack of direct correlation between the mean power frequency and fatigue level [5]. The validations of the above features are mainly based on comparing the time or workload of the thresholds in comparison to known critical fatigue thresholds, such as the ventilatory threshold, the neuromuscular threshold, or the lactate threshold (LT). However, few studies report methods to continuously monitor fatigue level in relation to lactate level or ventilatory oxygen uptake, to use as input into an intelligent system. One of the early attempts of an autonomous fatigue prediction system [6] used linear discriminant analysis to predict fatigue in a static exercise. Another study was successful in nding fatigue predicting features in a dynamic biceps curl exercise [7]. However, rarely are sporting activities static and localized to one muscle only, which is why vigorous activities, such as bicycling and running, are of interest to study further. Razanskas et al. [8] developed an algorithm using the cumulative sum of scores and correlation methods to train a machine-learning model to detect changes in lactate levels during bicycling. They found that a random forest (RF) model could predict the lactate level based on surface-electromyography (sEMG) signals with R 2 values>0.9 when trained using the time domain features and>0.8 when trained using the frequency domain features. However, for a model to be valid to use in a product application, it needs to be useful also for other types of activities, such as running. Furthermore, the model needs to be able to predict an individual's fatigue level without knowledge of the workload or the state of recovery. Furthermore, LT assessed by the absolute accumulation of lactate is a subjective variable [9,10], which may provide errors in determining the current fatigue state just by nding the lactate prediction level. Therefore, it may be of interest to further explore whether better prediction accuracy can be obtained using an a priori classi cation model that determines the fatigue level

recovery. Furthermore, LT assessed by the absolute accumulation of lactate is a subjective variable [9,10], which may provide errors in determining the current fatigue state just by nding the lactate prediction level. Therefore, it may be of interest to further explore whether better prediction accuracy can be obtained using an a priori classi cation model that determines the fatigue level based on lactate slope parameters. The purpose of this study was to investigate whether the machine learning method that was previously shown useful in detecting fatigue in bicycling using sEMG on lower limbs [8] can also predict fatigue in a running task. Furthermore, the study aimed to develop this method further by adding a segmentation of fatigue levels to the analysis in order to improve accuracy. 2. Materials and Methods The overall study design is described in a schematic diagram in Figure. In the rst step, data were acquired using a treadmill protocol, as described below, and sEMG sensors placed on the lower limb muscles of runners. Secondly, the sEMG signals were preprocessed. Three di erent types of feature sets were extracted from the signals, and a forward sequential feature elimination algorithm was used for feature selection. Two di erent RF models were trained, of which the rst model was used to predict lactate accumulation using regression, as described previously by Razanskas et al. [8], and the second model to classify fatigue based on change-point segmentation of lactate accumulation. The algorithm steps are described further below.

Sensors2019,19, 4729 3 of 17Sensors 2019, 19, x FOR PEER REVIEW 3 of 18 Figure 1. Schematic diagram of the study design. After data collection and data preprocessing, the study tested two ways to predict lactate levels of the athletes during the incremental running test, i.e., a novel fatigue classification model based on change-point segmentation and a previously described regression model for predicting lactate accumulation [8]. 2.1. Data Collection Twelve aerobically trained participants (five women and seven men) volunteered to participate in the study. Data were acquired during a single test session of an incremental treadmill-running test, as described below. Participants were either amateur long-distance runners (n = 4) or amateur triathlon athletes (n = 8) with a mean (±SD) running volume of 46.2 (±15.6) km/week, age 43 (±8) years, body mass 71.9 (±11.7) kg, and stature 1.75 (±0.99) m. All participants were informed about the study prior to the test occasion, both orally and in writing. Ethical approval was obtained from the regional ethical review board (Reg. No. 2014/162). Participants were equipped with wireless sEMG-sensors (Delsys Trigno, Delsys, Boston, MA, USA) on m. Vastus Lateralis (VL), m. Vastus Medialis (VM), m. Biceps femoris (BF), m. Semitendinosus (SM), and the medial head of m. Gastrocnemius (GM), as shown in Figure 2, abbreviated for the right leg as RVL, RVM, RBF, RSM, and RGM, respectively, and for the left leg as LVL, LVM, LBF, LSM, and LGM, respectively. Figure 1. Schematic diagram of the study design. After data collection and data preprocessing, the study tested two ways to predict lactate levels of the athletes during the incremental running test, i.e., a novel fatigue classi cation model based on change-point segmentation and a previously described regression model for predicting lactate accumulation [8]. 2.1. Data Collection Twelve aerobically trained participants ( ve women and seven men) volunteered to participate in the study. Data were acquired during a single test session of an incremental treadmill-running test, as described below. Participants were either amateur long-distance runners (n=4) or amateur triathlon athletes (n=8) with a mean ( SD) running volume of 46.2 ( 15.6) km/week, age

2.1. Data Collection Twelve aerobically trained participants ( ve women and seven men) volunteered to participate in the study. Data were acquired during a single test session of an incremental treadmill-running test, as described below. Participants were either amateur long-distance runners (n=4) or amateur triathlon athletes (n=8) with a mean ( SD) running volume of 46.2 ( 15.6) km/week, age 43 ( 8) years, body mass 71.9 ( 11.7) kg, and stature 1.75 ( 0.99) m. All participants were informed about the study prior to the test occasion, both orally and in writing. Ethical approval was obtained from the regional ethical review board (Reg. No. 2014/162). Participants were equipped with wireless sEMG-sensors (Delsys Trigno, Delsys, Boston, MA, USA) on m. Vastus Lateralis (VL), m. Vastus Medialis (VM), m. Biceps femoris (BF), m. Semitendinosus (SM), and the medial head of m. Gastrocnemius (GM), as shown in Figure, abbreviated for the right leg as RVL, RVM, RBF, RSM, and RGM, respectively, and for the left leg as LVL, LVM, LBF, LSM, and LGM, respectively. The sensor positions were palpated according to the SENIAM (Surface Electromyography for the Non-Invasive Assessment of Muscles) guidelines, and sites were shaved, rubbed, and cleaned before attachment of the sensors. The sensors were attached using double-sided tape (Trigno Sensor Skin Interface, Delsys, Boston, MA, USA), and elastic sports tape around the limb to keep the sensors securely xed. Maximum voluntary isometric contractions were performed for each muscle, according to the SENIAM guidelines, using manual resistance. Ventilatory measurements were performed using Oxycon Pro (Jaeger, Hoechberg, Germany), attached to a mask (7450 Series V2 Mask, Hans Rudolph Inc. Shawnee, Kansas, USA) worn by the participants to collect the ventilatory gases for analysis of ventilation, oxygen, and carbon dioxide. A heart rate monitor (Polar FT4, Kempele, Finland) was

Sensors2019,19, 4729 4 of 17 positioned on the thorax and connected to the Oxycon wirelessly. Subjective measures of leg fatigue and ventilatory fatigue were obtained using the Borg scale rating of perceived exertion, and blood lactate concentration samples (Lactate Pro 2, Arkray, Japan), were collected every two minutes throughout the test, by stepping to the side of the treadmill. This procedure took about 30 s. The blood samples were gathered from the right-hand ngertips of the participant. Prior to testing, each participant familiarized themselves with the perceived rating scale, the strategy to step aside from the treadmill to take a lactate sample, and the safety procedure when fatigue occurred. These procedures were practiced to enable safe and e cient data collection.Sensors 2019, 19, x FOR PEER REVIEW 4 of 18 Figure 2. The sensor positions used for surface electromyography. The sensor positions were palpated according to the SENIAM (Surface Electromyography for the Non-Invasive Assessment of Muscles) guidelines, and sites were shaved, rubbed, and cleaned before attachment of the sensors. The sensors were attached using double-sided tape (Trigno Sensor Skin Interface, Delsys, Boston, MA, USA), and elastic sports tape around the limb to keep the sensors securely fixed. Maximum voluntary isometric contractions were performed for each muscle, according to the SENIAM guidelines, using manual resistance. Ventilatory measurements were performed using Oxycon Pro (Jaeger, Hoechberg, Germany), attached to a mask (7450 Series V2 Mask, Hans Rudolph Inc. Shawnee, Kansas, USA) worn by the participants to collect the ventilatory gases for analysis of ventilation, oxygen, and carbon dioxide. A heart rate monitor (Polar FT4, Kempele, Finland) was positioned on the thorax and connected to the Oxycon wirelessly. Subjective measures of leg fatigue and ventilatory fatigue were obtained using the Borg scale rating of perceived exertion, and blood lactate concentration samples (Lactate Pro 2, Arkray, Japan), were collected every two minutes throughout the test, by stepping to the side of the treadmill. This procedure took about 30 s. The blood samples were gathered from the right-hand fingertips of the participant. Prior to testing, each participant familiarized themselves with the perceived rating scale, the strategy

perceived exertion, and blood lactate concentration samples (Lactate Pro 2, Arkray, Japan), were collected every two minutes throughout the test, by stepping to the side of the treadmill. This procedure took about 30 s. The blood samples were gathered from the right-hand fingertips of the participant. Prior to testing, each participant familiarized themselves with the perceived rating scale, the strategy to step aside from the treadmill to take a lactate sample, and the safety procedure when fatigue occurred. These procedures were practiced to enable safe and efficient data collection. The incremental treadmill protocol started with a light warm-up at 8 (for women) or 9 (for men) km/h. The test started with six minutes of running at a speed corresponding to just below each athlete’s 10 km tempo, which was between 11 and 13 km/h. After the initial six minutes, an increase in workload (1–1.5 km/h) was performed every two minutes until the athlete reached a lactate level above 5 mmol/l (or was unable to increase the running workload). At this stage, the workload level was kept constant until the athlete was unable to continue (minimum six minutes). The final six minutes of the test were at the initial speed, i.e., their individual 10 km tempo. 2.2. Feature Extraction The sEMG signals were recorded from the muscles at a sampling rate of 1926 Hz. In order to eliminate electronic noise and motion artifacts, the sampled signals were preprocessed using the suggested Butterworth filter [11] with 10th order 400 Hz low pass filter at a stopband of 450 Hz with Figure 2.The sensor positions used for surface electromyography. The incremental treadmill protocol started with a light warm-up at 8 (for women) or 9 (for men) km/h. The test started with six minutes of running at a speed corresponding to just below each athlete's 10 km tempo, which was between 11 and 13 km/h. After the initial six minutes, an increase in workload (1–1.5 km/h) was performed every two minutes until the athlete reached a lactate level above 5 mmol/L (or was unable to increase the running workload). At this stage, the workload

of running at a speed corresponding to just below each athlete's 10 km tempo, which was between 11 and 13 km/h. After the initial six minutes, an increase in workload (1–1.5 km/h) was performed every two minutes until the athlete reached a lactate level above 5 mmol/L (or was unable to increase the running workload). At this stage, the workload level was kept constant until the athlete was unable to continue (minimum six minutes). The nal six minutes of the test were at the initial speed, i.e., their individual 10 km tempo. 2.2. Feature Extraction The sEMG signals were recorded from the muscles at a sampling rate of 1926 Hz. In order to eliminate electronic noise and motion artifacts, the sampled signals were preprocessed using the suggested Butterworth lter [11] with 10th order 400 Hz low pass lter at a stopband of 450 Hz with 60 dB attenuation and the 10th order 20 Hz high pass lter at a stopband of 10 Hz with 60 dB attenuation. The ltered signal S(t) was interpolated using Hermite cubic splines and used for feature extraction and analysis. 2.2.1. Frequency Domain Features In order to extract frequency domain features, the preprocessed signal S(t) was segmented into muscle activity bursts Si(t) using a method for detecting muscle activity previously described by Razanskas et al. [8]. An i th muscle activity burst corresponds to a single stride cycle in the running. The power spectrum Pi(f) of an i th burst was computed by using discrete Fourier transform on Si(t) using Equation (1):

Sensors2019,19, 4729 5 of 17 Pi(f)= DFT(Si) . (1) Pi(f) was further used to compute the power distribution of frequencies Di(f) of each i th burst using Equation (2): Di(f)= Pi(f) R f Nyq 0 Pi(f)df , (2) where fNyqis the Nyquist frequency. Frequency-based features were extracted using Si(t) and Di(f). The description of each feature is given in Table. Details can be found elsewhere [12]. Table 1. Frequency-based features extracted from an i th segment of the surface-electromyography (sEMG) signal S i(t) obtained from each muscle. Number Name Description 1 RMS Root mean square error of S i(t). 2 dRMS The backward di erence of the root mean square error of S i(t). 3 IF Instantaneous frequency or zero-crossings of S i(t) divided by two. 4 ModF Mode of D i(f). 5 MnF Mean of D i(f). 6 StD The standard deviation of D i(f). 7 Skew The skewness of D i(f). 8 Kurt Kurtosis of D i(f). 9–17 q 0.1to q 0.9 Every 10th percentile of D i(f), i.e., q 0.5is the 50th percentile or the median power frequency. 18–36 p 23-47 Hzto p 234-258 Hz Relative power contained in 23.44 Hz bands (width of 6 DFT bins) of power spectrum P i(f) with an overlap of 11.72 Hz, starting with a band of 23.44–46.88 Hz and ending with a band of 234.4–257.8 Hz. 2.2.2. Time-Domain Features The muscle activation time is a critical time-domain feature that is de ned as the time di erence between the activation and deactivation of a muscle during a stride. The activation moment A(i) and the deactivation moment D(i) of a muscle were computed using the local maxima and the local minima of the derivative of the sEMG signal S(t), as described by Razanskas et al. [12]. The time span of a single stride cycle was estimated using two consecutive muscle activation moments, A(i) and A(i+1). It was observed that muscles BF, VM, and VL red sequentially during a stride cycle, i.e., the activation moment of BF preceded the activation moment of VM, and the activation moment of VM preceded the activation

by Razanskas et al. [12]. The time span of a single stride cycle was estimated using two consecutive muscle activation moments, A(i) and A(i+1). It was observed that muscles BF, VM, and VL red sequentially during a stride cycle, i.e., the activation moment of BF preceded the activation moment of VM, and the activation moment of VM preceded the activation moment of VL. Hence, the length of time between A(i+1) and A(i) of muscle BF corresponds to the time span of one stride cycle, and the activation of other muscles can be computed as fractions of this baseline time length. We used the activation moment A(i) of RBF as the starting timestamp because it was the rst activation signal to start each activation cycle. It is pertinent that Razanskas et al. [8,12] used three sEMG channels recorded from rectus femoris, vastus lateralis, and semitendinosus muscles of both legs for estimating bicycling fatigue. In order to develop a similar model for estimating running fatigue, we used three sEMG channels recorded from BF, VM, and VL, and discarded SM and GM when extracting features based on the time domain analysis. The phase shift ØX, Ybetween the activation moments of two muscles, say X and Y, in one stride cycle, was computed using Equation (3): ?X, Y(i)= AY(i) Ax(i) ARBF(i+1) ARBF(i) . (3) Similarly, the active time percentage Xof muscle X in one stride cycle was computed using Equation (4): X(i)= Dx(i) Ax(i) ARBF(i+1) ARBF(i) . (4)

Sensors2019,19, 4729 6 of 17 The root-mean-square Xof the sEMG signal S(t) of muscle X for the i th cycle was computed using Equation (5): X(i)= 1 Dx(i) Ax(i) . (5) Furthermore, the arithmetic mean (AM) and standard deviation (SD) of features ØX, Y, X,and X were computed using Equations (6) and (7), respectively: AM= 1 n nX i=1 ai, (6) SD= vt 1 n 1 nX i=1 (ai a) 2 , (7) whereaiis the feature value for ani th stride cycle, andnis the total stride cycles. Moreover, the asymmetry of each feature was measured by computing the corresponding arithmetic means of that feature from the right and the left leg and nding the absolute di erence between them, given as: X ( )= Mean Right( ) Mean le f t( ) . (8) In order to reduce the model complexity, Razanskas et al. [8] used a subset by excluding amplitude-based features from the time domain features. The subset was termed as the time event features. The time-domain and time event features (shaded grey) are listed in Table. Table 2. Time domain and time event features (shaded grey) of sEMG signals. Two numbers assigned to a feature correspond to the arithmetic mean and standard deviation of that feature. Number Symbol Number Symbol The mean and standard deviation of phase shifts? X,Ybetween X and Y muscles for total strides The mean and standard deviation of the signal root-mean-square Xof muscle X for total strides 1,2 ? RBF, RVM 31, 32 RBF 3,4 ? RBF, RVL 33, 34 RVM 5,6 ? RVM, RVL 35, 36 RVL 7,8 ? LBF, LVM 37, 38 LBF 9,10 ? LBF, LVL 39, 40 LVM 11,12 ? LVM, LVL 41, 42 LVL 13,14 ? RBF,LBF Asymmetry of? X,Ybetween the right and left leg. 15,16 ? RVM, LVM 43 P (?BF,VM) 17,18 ? RVL, LVL 44 P (?BF, VL) The mean and standard deviation of the active time percentages Xof muscle X for total strides 45 P (?VM, VL) 19,20 RBF Asymmetry of Xbetween the right and left leg. 21,22 RVM 46 P BF 23, 24 RVL 47

The study investigates a method to classify running fatigue using surface electromyography.