Abstract
tigue is a multifunctional and complex phenomenon that a ects how individuals perform an activity. Fatigue during running causes changes in normal gait parameters and increases the risk of injury. To address this problem, wearable sensors have been proposed as an unobtrusive and portable system to measure changes in human movement as a result of fatigue. Recently, a category of wearable devices that has gained attention is exible textile strain sensors because of their ability to be woven into garments to measure kinematics. This study uses exible textile strain sensors to continuously monitor the kinematics during running and uses a machine learning approach to estimate the level of fatigue during running. Five female participants used the sensor-instrumented garment while running to a state of fatigue. In addition to the kinematic data from the exible textile strain sensors, the perceived level of exertion was monitored for each participant as an indication of their actual fatigue level. A stacked random forest machine learning model was used to estimate the perceived exertion levels from the kinematic data. The machine learning algorithm obtained a root mean squared value of 0.06 and a coe cient of determination of 0.96 in participant-speci c scenarios. This study highlights the potential of exible textile strain sensors to objectively estimate the level of fatigue during running by detecting slight perturbations in lower extremity kinematics. Future iterations of this technology may lead to real-time biofeedback applications that could reduce the risk of running-related overuse injuries. Keywords:fatigue running; kinematics; soft sensors; random forest 1. Introduction Running is one
scenarios. This study highlights the potential of exible textile strain sensors to objectively estimate the level of fatigue during running by detecting slight perturbations in lower extremity kinematics. Future iterations of this technology may lead to real-time biofeedback applications that could reduce the risk of running-related overuse injuries. Keywords:fatigue running; kinematics; soft sensors; random forest 1. Introduction Running is one of the most popular and healthy activities [1,2] worldwide but carries a high risk of injury [3]. Long-distance running is, by nature, a prolonged and repetitive activity, which can induce fatigue. Fatigue accumulates as runners increase distance or intensity and is de ned as an exercise-induced reduction in the ability to generate muscle force or power that is caused by changes in the neural drive or exhaustion of muscle contractile function [4]. Fatigue is commonly measured either by direct physiological means (e.g., heart rate, blood lactate concentration, etc.) or by subjective rating of exertion. Perceived rating of exertion is a subjective indication of fatigue that integrates information from the peripheral muscles, the central neural system, and the cardiovascular system [5]. The Borg Rating of Perceived Exertion (RPE) scale [6] is a popular and practical tool that has been widely used in running research [79]. RPE scores are strongly correlated with blood lactate concentration and can be used both in the laboratory and in Sensors2020,20, 5573; doi:10.3390 /s20195573 /journal/sensors
Sensors2020,20, 5573 2 of 11 the eld [1012]. A recent systematic review reported that RPE is a more sensitive measure of acute stress (fatigue) than objective measures, such as blood lactate concentration and heart rate [13], as it encompasses both the psychological and physiological components of fatigue [14]. Rating of Perceived Exertion has been increasingly used in running injury research, as well as in clinical contexts, because of its important contribution to the overall calculation of training loadand, therefore, injury riskfor an individual [15,16]. Fatigue can a ect both performance and injury risk, manifested by changes in lower extremity biomechanicsparticularly joint kinematics. Changes in kinematics and the shift in work from distal to proximal joints [17] is believed to reduce running economy during runs to exhaustion [18,19]. Running in a fatigued state also exposes the body to greater mechanical forces, which may increase risk of injury [2022]. The greater loads experienced by the lower extremities when fatigued are due to changes in kinematics at initial contact and during the early loading phase of running [2326]. Kinematic [20,26,27], kinetic [28,29], and electromyographic (EMG) [3032] changes may occur in response to fatigue. Traditionally, these changes have been measured by optical motion capture systems, force plate or accelerometer analysis, and surface EMG (sEMG), respectively. While kinetics and EMG can be measured remotely (i.e., outside of a lab environment), the remote measurement of kinematics is more challenging. The concept that increasing fatigue during a prolonged run is correlated with a change in kinematics is not new [20,26,27]. Assuming that the kinematic changes are as predictable (and progressive) as the level of fatigue, the fatigue level of a runner could be estimated by continuously monitoring their kinematics using a wearable device. From a machine learning perspective, the amount of information gathered from lower extremity kinematics may yield a better estimation of fatigue than monitoring kinetics or EMG activity. As kinematic changes associated with fatigue may decrease performance and increase injury risk [33], the ability to monitor changes in kinematics over the course of a run, especially outside of a lab or clinic setting, has substantial
perspective, the amount of information gathered from lower extremity kinematics may yield a better estimation of fatigue than monitoring kinetics or EMG activity. As kinematic changes associated with fatigue may decrease performance and increase injury risk [33], the ability to monitor changes in kinematics over the course of a run, especially outside of a lab or clinic setting, has substantial implications on both performance and injury risk. Two main methods have been used to measure changes in kinematics to detect fatigue: (1) optical motion capture and (2) wearable devices. Wearable sensors have the advantage of unobtrusive measurement in ecologically valid environments during daily life. Inertial measurement units (IMUs) are the primary wearable device that have been used to detect the state of fatigue or to quantitatively measure fatigue level. A machine learning approach has been used to estimate the level of fatigue based on information from wearable sensors, including Strohrmann et al.'s [34] investigation into the correlation of features extracted from IMUs with the level of fatigue during a prolonged run. There was a strong correlation between IMU signal changes from 12 IMUs mounted on the lower and upper body and the perceived level of exertion of runners. Other studies have focused on the binary detection of the state of fatigue and non-fatigue based on measurements by IMUs [35,36]. Tibia-mounted IMUs have obtained a high classi cation accuracy of fatigue during running and occupational tasks [36]. Karg et al. [37] used a hidden Markov model as a regression model to predict fatigue level during the performance of a squat exercise. This study monitored the kinematics of the lower and upper body using an optical motion capture system and obtained a high accuracy level. To our knowledge, the use of exible textile (soft) strain sensors for continuous monitoring of fatigue has not yet been studied. Flexible textile strain sensors are a category of wearable sensors that have recently been used for human gait analysis [3840], trunk motion monitoring [41,42], and hand gesture recognition [43]. These sensors work by measuring changes in the resistance or capacitance when they are elongated [44]. The
textile (soft) strain sensors for continuous monitoring of fatigue has not yet been studied. Flexible textile strain sensors are a category of wearable sensors that have recently been used for human gait analysis [3840], trunk motion monitoring [41,42], and hand gesture recognition [43]. These sensors work by measuring changes in the resistance or capacitance when they are elongated [44]. The advantages of exible textile strain sensors are their easy integration into clothing and convenient use compared to rigid IMUs. In a previous study, we optimized a wearable exible textile strain sensor system for three dimensional lower extremity joint angle measurement and subsequently validated this system for the measurement of lower extremity running kinematics [38]. In this study, we extend our previous research to use exible textile strain sensors for the estimation of fatigue level during running. Our aim was to estimate the level of fatigue (measured by Borg's RPE scale) during the course of a prolonged run based on the lower extremity kinematic information
Sensors2020,20, 5573 3 of 11 provided by the exible textile strain sensors. We hypothesized that by using information from the exible textile strain sensors and using a machine learning approach, we would be able to accurately estimate the level of fatigue at a given moment for an individual during a prolonged run. 2. Materials and Methods 2.1. Experiment Setup A resistance-based exible textile strain sensor was employed in this study. The strain sensors were made of spandex multi lament yarn coated with a carbon black thermoplastic elastomer composite [41]. The sensors were conditioned with sinusoidal strain of 40% at 10% per second before use. The sensor showed highly linear performance and no hysteresis in the working range of up to 30% strain. The sensor was then coated with an insulating sheath to prevent shortening of the circuit if participants sweat during the prolonged run. The sensor characteristics are summarized in Table. Table 1.Characteristics of the strain sensor. Characteristics Condition Linearity Highly linear <30% strain Hysteresis Zero for <30% strain Frequency response up to 10 Hz Gauge factor 5 Signal drift No drift up to 4 h The resistance-based strain sensor's fundamental work is expressed in Equation (1), where the resistance value is a ected by the increase in the sensor's length: DR= DL L0 (1) The sensor placement was optimized in a previous study [38] and four positions were selected by genetic algorithm for a multi-axis hip joint. The positions of the sensors on the knee and ankle were selected according to the joint axes. The wearable system was validated for lower extremity joint angle measurement during running and showed an error of less than 2.5 for multi-axis lower body kinematic monitoring during running [38]. The prototype had a total of nine sensors: four sensors at the hip, two sensors at the knee, and three sensors at the ankle. However, one of the knee sensors was excluded from data processing because of high noise. The circuit used for data recording was a voltage divider circuit with a resistance of 40 kW. Participants were asked to run on a split-belt treadmill
of nine sensors: four sensors at the hip, two sensors at the knee, and three sensors at the ankle. However, one of the knee sensors was excluded from data processing because of high noise. The circuit used for data recording was a voltage divider circuit with a resistance of 40 kW. Participants were asked to run on a split-belt treadmill (Bertec Corporation, Columbus, OH, USA) while the strain sensor signals were recorded at a frequency of 100 Hz by a data acquisition device. The prototype and the schematic of data processing are outlined in Figure. The raw signals of the sensor did not drift over the course of 45 min of running. Similarly, the participant's sweat did not a ect the performance of the sensor. 2.2. Data Collection Five healthy, pain-free female recreational runners participated in this study. The selection of female participants was due to known di erences in kinematics between male and female runners [45,46]. To obtain a homogenous sample, participants were selected based on running experience and recent performance in a 10-kilometer race. Participants were also screened to ensure adequate t of the prototype garment. Onset of the COVID-19 pandemic and the restrictions on in-person research limited data collection to a sample size of ve. Participants ran in standardized running shoes (New Balance 880v10, Boston, MA, USA) to ensure that biomechanical changes were not due to di erences in shoe characteristics. The use of the Borg RPE scale was explained to the participants prior to the experiment. For consistency, participants were
Sensors2020,20, 5573 4 of 11 asked to report the conscious sensation of how hard you are driving your working limbs and how strenuous the exercise feels at this point in time. After putting on the instrumented running tights, participants were given three minutes to warm up and become familiar with the treadmill, running at a steady speed of 8 km/h. The treadmill was then stopped, and participants were asked to rate their current level of exertion on the Borg RPE scale before commencing the test. Data collection began with participants running at 8 km/h. Participants were asked to report their RPE score every three minutes. If the RPE score was lower than 13 (somewhat hard), the speed was increased by 1 km/h. Once the participant reached an RPE score of 13, the speed was kept constant and they were asked to continue to report their RPE score every three minutes until they reported a level of 17 (very hard). At this point, one more minute of data was recorded before terminating the test. FigureB shows the individual changes in the participants' RPE scores during the test.Sensors 2020, 20, x 4 of 12 Figure 1. Strain sensor positions on a pair of running tights and the data processing schematic. DAQ, data acquisition device; ML, machine learning. 2.2. Data Collection Five healthy, pain-free female recreational runners participated in this study. The selection of female participants was due to known differences in kinematics between male and female runners [45,46]. To obtain a homogenous sample, participants were selected based on running experience and recent performance in a 10-kilometer race. Participants were also screened to ensure adequate fit of the prototype garment. Onset of the COVID-19 pandemic and the restrictions on in-person research limited data collection to a sample size of five. Participants ran in standardized running shoes (New Balance 880v10, Boston, MA, USA) to ensure that biomechanical changes were not due to differences in shoe characteristics. The use of the Borg RPE scale was explained to the participants prior to the experiment. For consistency, participants were asked to report “the conscious sensation of how
collection to a sample size of five. Participants ran in standardized running shoes (New Balance 880v10, Boston, MA, USA) to ensure that biomechanical changes were not due to differences in shoe characteristics. The use of the Borg RPE scale was explained to the participants prior to the experiment. For consistency, participants were asked to report “the conscious sensation of how hard you are driving your working limbs and how strenuous the exercise feels at this point in time.” After putting on the instrumented running tights, participants were given three minutes to warm up and become familiar with the treadmill, running at a steady speed of 8 km/h. The treadmill was then stopped, and participants were asked to rate their current level of exertion on the Borg RPE scale before commencing the test. Data collection began with participants running at 8 km/h. Participants were asked to report their RPE score every three minutes. If the RPE score was lower than 13 (“somewhat hard”), the speed was increased by 1 km/h. Once the participant reached an RPE score of 13, the speed was kept constant and they were asked to continue to report their RPE score every three minutes until they reported a level of 17 (“very hard”). At this point, one more minute of data was recorded before terminating the test. Figure 2B shows the individual changes in the participants’ RPE scores during the test. Staked ML Algorithm DAQ Segmentation 3 sensors on anterior hip 1 sensor on posterior hip 1 sensor on anterior knee 2 sensors on anterior ankle 1 sensor on posterior kl Feature Extraction Level Description 6 None 7-12 Very light- light 13-14 Somewhat hard 15-16 Hard 17-18 Very hard 19-20 Very, very hard A B Figure 1. Strain sensor positions on a pair of running tights and the data processing schematic. DAQ, data acquisition device; ML, machine learning.Sensors 2020, 20, x 4 of 12 Figure 1. Strain sensor positions on a pair of running tights and the data processing schematic. DAQ, data acquisition device; ML, machine learning. 2.2. Data Collection Five healthy, pain-free female recreational runners participated
Strain sensor positions on a pair of running tights and the data processing schematic. DAQ, data acquisition device; ML, machine learning.Sensors 2020, 20, x 4 of 12 Figure 1. Strain sensor positions on a pair of running tights and the data processing schematic. DAQ, data acquisition device; ML, machine learning. 2.2. Data Collection Five healthy, pain-free female recreational runners participated in this study. The selection of female participants was due to known differences in kinematics between male and female runners [45,46]. To obtain a homogenous sample, participants were selected based on running experience and recent performance in a 10-kilometer race. Participants were also screened to ensure adequate fit of the prototype garment. Onset of the COVID-19 pandemic and the restrictions on in-person research limited data collection to a sample size of five. Participants ran in standardized running shoes (New Balance 880v10, Boston, MA, USA) to ensure that biomechanical changes were not due to differences in shoe characteristics. The use of the Borg RPE scale was explained to the participants prior to the experiment. For consistency, participants were asked to report “the conscious sensation of how hard you are driving your working limbs and how strenuous the exercise feels at this point in time.” After putting on the instrumented running tights, participants were given three minutes to warm up and become familiar with the treadmill, running at a steady speed of 8 km/h. The treadmill was then stopped, and participants were asked to rate their current level of exertion on the Borg RPE scale before commencing the test. Data collection began with participants running at 8 km/h. Participants were asked to report their RPE score every three minutes. If the RPE score was lower than 13 (“somewhat hard”), the speed was increased by 1 km/h. Once the participant reached an RPE score of 13, the speed was kept constant and they were asked to continue to report their RPE score every three minutes until they reported a level of 17 (“very hard”). At this point, one more minute of data was recorded before terminating the test. Figure 2B shows the individual
increased by 1 km/h. Once the participant reached an RPE score of 13, the speed was kept constant and they were asked to continue to report their RPE score every three minutes until they reported a level of 17 (“very hard”). At this point, one more minute of data was recorded before terminating the test. Figure 2B shows the individual changes in the participants’ RPE scores during the test. Staked ML Algorithm DAQ Segmentation 3 sensors on anterior hip 1 sensor on posterior hip 1 sensor on anterior knee 2 sensors on anterior ankle 1 sensor on posterior kl Feature Extraction Level Description 6 None 7-12 Very light- light 13-14 Somewhat hard 15-16 Hard 17-18 Very hard 19-20 Very, very hard A B Figure 2. (A) Description of levels of Borg Rating of Perceived Exertion (RPE) scale and (B) changes in the Borg RPE of each participant during the test. 2.3. Data Segmentation Two di erent methods were used to segment the strain sensor signals: (1) based on strides and (2) based on a moving window over the data. In method (1), the peak values of the rst sensor were detected and used to determine strides. In method (2), the peak values of the rst signal were detected
Sensors2020,20, 5573 5 of 11 and a window of time prior to this data point was selected. Window lengths of 0.5, 1, 1.5, and 2 seconds were examined. The window length of 1 second obtained the best results (Figure).Sensors 2020, 20, x 5 of 12 Figure 2. (A) Description of levels of Borg Rating of Perceived Exertion (RPE) scale and (B) changes in the Borg RPE of each participant during the test. 2.3. Data Segmentation Two different methods were used to segment the strain sensor signals: (1) based on strides and (2) based on a moving window over the data. In method (1), the peak values of the first sensor were detected and used to determine strides. In method (2), the peak values of the first signal were detected and a window of time prior to this data point was selected. Window lengths of 0.5, 1, 1.5, and 2 seconds were examined. The window length of 1 second obtained the best results (Figure 3). Figure 3. The window-based and stride-based approached for sensor data segmentation. 2.4. Feature Extraction A set of statistical and temporal features were extracted according to the biomechanical changes of the lower extremity associated with fatigue. The features were: mean; minimum; maximum; range of motion; stride length—in addition, variation of mean; minimum; maximum; range of motion during the last 10 strides/windows. Moreover, the strides/windows were segmented to five sub- segments and the mean, minimum, and maximum values of each sub-segment were extracted. The same features were extracted from the first derivative of the strain sensor’s signals. 2.5. Machine Learning Algorithm Random forest is an ensemble of decision trees that has shown promising results when compared with conventional machine learning algorithms including support vector machine and neural networks in regression and classification applications for strain sensor’s data analysis [39,47]. Random forest models are robust to outliers, nonlinear and unbalanced data, and produce low bias and moderate variance [48,49]. We have previously used random forest to accurately estimate joint angles using strain sensors [39]. We therefore chose random forest as our method for data analysis. Deep learning models were
Description
The study explores wearable sensors for monitoring fatigue in running.