Comparison between Accelerometer and Gyroscope in Predicting Level-Ground Running Kinematics by Treadmill Running Kinematics Using a Single Wearable Sensor

earable sensors facilitate running kinematics analysis of joint kinematics in real running environments. The use of a few sensors or, ideally, a single inertial measurement unit (IMU) is preferable for accurate gait analysis. This study aimed to use a convolutional neural network (CNN) to predict level-ground running kinematics (measured by four IMUs on the lower extremities) by using treadmill running kinematics training data measured using a single IMU on the anteromedial side of the right tibia and to compare the performance of level-ground running kinematics predic- tions between raw accelerometer and gyroscope data. The CNN model performed regression for intraparticipant and interparticipant scenarios and predicted running kinematics. Ten recreational runners were recruited. Accelerometer and gyroscope data were collected. Intraparticipant and interparticipant R 2 values of actual and predicted running

the anteromedial side of the right tibia and to compare the performance of level-ground running kinematics predic- tions between raw accelerometer and gyroscope data. The CNN model performed regression for intraparticipant and interparticipant scenarios and predicted running kinematics. Ten recreational runners were recruited. Accelerometer and gyroscope data were collected. Intraparticipant and interparticipant R 2 values of actual and predicted running kinematics ranged from 0.85 to 0.96 and from 0.7 to 0.92, respectively. Normalized root mean squared error values of actual and predicted running kinematics ranged from 3.6% to 10.8% and from 7.4% to 10.8% in intraparticipant and interparticipant tests, respectively. Kinematics predictions in the sagittal plane were found to be better for the knee joint than for the hip joint, and predictions using the gyroscope as the regressor were demonstrated to be signi cantly better than those using the accelerometer as the regressor. Keywords: deep learning; convolutional neural network; running; kinematics; wearable sensor; running kinematics analysis; accelerometer; gyroscope 1. Introduction Although running is a popular recreational sport worldwide, injury incidence rates ranging from 19.4% to 79.3% have been reported [1]. These injuries are mostly due to overuse of the lower extremities. Among various lower extremity injuries, injuries in the region surrounding the knee were found to be the most common and have severe consequences [1], such as patellofemoral pain syndrome, anterior cruciate ligament in- jury, iliotibial band friction syndrome, and patellar tendinopathy [2,3]. Therefore, the investigation of lower extremity kinematics is important for the better understanding of running-related injuries. In human running kinematics analysis, a camera-based motion capture system is considered the gold standard [4]. The capture system is constructed in a laboratory. The research participant is surrounded by cameras and retrore ective markers are placed on anatomical landmarks of the human body [4]. Despite the use of multiple cameras, a body part or object can occlude the eld of view between a marker and the cameras, decreasing Sensors2021,21, 4633.

the human body [4]. Despite the use of multiple cameras, a body part or object can occlude the eld of view between a marker and the cameras, decreasing Sensors2021,21, 4633.

Sensors2021,21, 4633 2 of 13 the effectiveness of such motion tracking systems. In addition, despite the valid and reliable measurements provided by such a system, the technology is con ned to the laboratory setting and requires expensive, sophisticated equipment that is permanently installed in a large room [5–7]. Therefore, although camera-based motion capture systems are the gold standard, making valid assessments of level-ground running kinematics using these systems is still challenging. With technological advancements in sensors and data analysis, real-time mobile sensor devices have been developed. Wearable, mobile sensors can be used to analyze human gait patterns in the real world and in outdoor running environments [8]. Inertial measurement units (IMUs)—including accelerometers, gyroscopes, and magnetometers—have been employed extensively in recent studies to detect acceleration and angular velocity [9–13]. Human gait kinematics, such as that pertaining to joint angles, can be estimated by merging inertial data from two body segments [14–16]. Compared with camera-based motion capture systems, wearable sensors are more portable and can monitor limb kinematics in the eld without obstructing normal gait [5]. Numerous studies have examined the validity and reliability of using wearable sensors for joint angle estimation during level walking [17–20] and running [18,21–23]. Typically, joint angle assessment requires at least two IMU sensors, with each placed on one body segment of the joint studied [24–26]. Therefore, multiple sensors are required. Usually, 18 sensors are used in total, with 11 sensors placed on the upper body and 7 placed on the lower extremities [16,27–31]. To reduce redundancy and cost, a study attempted to minimize the number of sensors used during human kinematics analysis [32]. Two approaches have been proposed to re- duce the number of sensors, namely, the model-based approach [33–35] and the data-driven approach [5,36,37]. The model-based approach involves establishing a serial kinematic chain model and calculating inverse kinematics to estimate the motion of the lower extrem- ities [5,35]. Several studies have investigated lower extremities kinematics with reduced numbers of IMU sensors [33–35]. A study has used a single IMU sensor to detect age- and surface-related differences in walking with machine learning algorithms [38]. However, model-based

model-based approach involves establishing a serial kinematic chain model and calculating inverse kinematics to estimate the motion of the lower extrem- ities [5,35]. Several studies have investigated lower extremities kinematics with reduced numbers of IMU sensors [33–35]. A study has used a single IMU sensor to detect age- and surface-related differences in walking with machine learning algorithms [38]. However, model-based approaches have also been criticized for errors arising from the misalignment of the sensors during the set-up or the trial (see [5,39]). By contrast, the data-driven approach employs supervised machine learning models to estimate lower extremity kinematics [5,36]. Numerous studies have applied the machine learning method to predict extremities kinematics during running orwalking [ . Zimmermann et al. [26] determined the time dynamic features of lower body kinematics for improved orientation alignment and IMU-to-segment assignment tasks by using a deep learning approach that included CNNs combined with long short-term memory networks and generalized recurrent units. Lastly, Gholami et al. [5] developed a novel method using a single shoe-mounted accelerometer and CNN to estimate lower extremity gait kinematics in the sagittal plane during treadmill running. These data-driven approaches minimize the number of IMUs while maximizing the accuracy of the measurements by using supervised machine learning models. Different types of IMUs, namely accelerometers and gyroscopes, have been used to estimate human gait kinematics with differing performance [14–16]. Rhudy and Ma- honey [41] reported that estimations in step counting were better when using gyroscopic sensors than when using accelerometer sensors. Mahoney and Rhudy [42] also presented a machine learning method in gait stride categorization (i.e., walking, jogging, or running). Arti cial neural network models trained with raw accelerometer data performed better (speci cally, categorizing gait stride more accurately) than those trained using gyroscopic data (see [42]). These results demonstrate that different sensors have different advan- tages for encoding level-ground gait characteristics. However, performance in predicting running kinematics using accelerometers or gyroscopes is uncon rmed. These studies [5,8,10–12,15–24,26–40] have pertained to the use of camera-based mo- tion analysis systems to capture target kinematics. The myoMotion (Noraxon, Scottsdale,

data (see [42]). These results demonstrate that different sensors have different advan- tages for encoding level-ground gait characteristics. However, performance in predicting running kinematics using accelerometers or gyroscopes is uncon rmed. These studies [5,8,10–12,15–24,26–40] have pertained to the use of camera-based mo- tion analysis systems to capture target kinematics. The myoMotion (Noraxon, Scottsdale,

Sensors2021,21, 4633 3 of 13 USA) sensors are a set of sensors for speci c body locations (i.e., 7 sensors for the lower extremities or 16 sensors for the full body), whereas IMeasureU (Vicon, Oxford, UK) sensors can be used individually and independently to capture raw data at a given location. The attachment of the IMeasureU sensor at the anteromedial tibia is also more convenient for users. Because IMUs have been shown to be reliable and accurate for use outside of laboratory settings, recent studies have aimed to reduce the number of sensors worn, with the ultimate aim of using only a single IMU, to measure running kinematics. This study used two sets of IMUs: one (myoMotion) was used to capture target joint kinematics, whereas another one (IMeasureU) was used to capture regressor data for modeling the target joint kinematics using a deep learning approach. This study aimed to (1) use a CNN model to predict level-ground running kinematics, measured by four IMUs on the lower extremities of the right side, using treadmill running kinematics training data, measured using a single IMU on the anteromedial side of the right tibia and (2) compare the perfor- mance of level-ground running kinematics predictions between raw accelerometer and gyroscope data. 2. Materials and Methods 2.1. Participants Five male and ve female healthy recreational runners (age: 22.70 1.34 years, height: 168 6.32 cm, and weight: 61.33 6.82 kg) were recruited to participate in this study. The participants had no history of injury within the previous 6 months. Written informed consent (approved by the Human Research Ethics Committee of the University) was obtained from each participant before data collection. 2.2. Instruments All participants were equipped with the two sets of IMUs. Speci cally, myoMotion sensors were used to capture lower limb movement. According to the sensor placement for lower extremities recommended by the manufacturer (Noraxon), seven myoMotion sensors were attached to the participants with either an elastic strap or belt on the pelvis, left and right thighs, left and right shanks, and left and right feet. The sensors' placement is indicated in Figure. Only data

were used to capture lower limb movement. According to the sensor placement for lower extremities recommended by the manufacturer (Noraxon), seven myoMotion sensors were attached to the participants with either an elastic strap or belt on the pelvis, left and right thighs, left and right shanks, and left and right feet. The sensors' placement is indicated in Figure. Only data collected by the four sensors attached on the pelvis and right leg were considered and used for data analysis. Accelerations (on thex,y, andzaxes) and anatomical angles for the lower limbs (knee exion/extension and hip exion/extension) were measured with a sampling rate of 100 Hz. Second, a single IMeasureU sensor was attached to the anteromedial side of the right tibia, and tri-axial accelerations and angular velocities were recorded. The raw data of the accelerometer and gyroscope were acquired, at a sampling rate of 500 Hz, and used for data processing. The two sets of sensors were synchronized with three vertical right foot strikes at the beginning of each trial [40]. The directions of the axes of IMeasureU and myoMotion are indicated in Appendix, respectively.Sensors 2021, 21, x FOR PEER REVIEW 4 of 14 Figure 1. The sensors’ placement on the participant. 2.3. Experiment Protocol and Data Collection All participants were required to complete the Physical Activity Readiness Question- naire (PAR-Q, Canadian Society for Exercise Physiology, 2002. www.csep.ca/forms, 30 September 2020) and a medical history questionnaire before the study began. Those with a “Yes” to one or more questions in the PAR-Q, or with any obvious anatomical abnor- malities, were excluded. Anthropometric data, including height, weight, and age, were measured and collected. All running trials were conducted on a treadmill in the labora- tory. Participants were advised to wear their usual running attire with self-selected run- ning shoes and that they should not participate in any physical activity on the day prior to testing. The IMeasureU and myoMotion sensors were equipped, calibrated, and syn- chronized. After synchronization, participants warmed up at 1.5 m/s on the treadmill for 30 s. Every 3 min, the speed of the treadmill was increased, first to 2.0

attire with self-selected run- ning shoes and that they should not participate in any physical activity on the day prior to testing. The IMeasureU and myoMotion sensors were equipped, calibrated, and syn- chronized. After synchronization, participants warmed up at 1.5 m/s on the treadmill for 30 s. Every 3 min, the speed of the treadmill was increased, first to 2.0 m/s, then to 2.5 m/s, to 3.0 m/s, and finally to 3.5 m/s [5,43]. The participants were then asked to run in an indoor squash court for 3 min at their preferred speed (average speed: 2.44 ± 0.34 m/s), which was recorded using a Brower Timing System (Brower Timing Systems, Draper, UT, USA). The data obtained from the two sets of IMUs were compared. 2.4. The Deep Learning Regression Model The CNN deep learning model used by Gholami et al. [5] was used to compare the results. Gholami et al. [5] used four layers of one-dimensional convolutional layers (Conv1D) to model the target variable array Y containing the hip, knee, and ankle flexion values. A four-column array X was constructed with the first three columns correspond- ing to the orthogonal accelerations (a x, ay, and az) and the last column corresponding to the magnitude of the total accelerations, a xyz. The rows of X correspond to the IMU data collected at 100 Hz. X was separated into overlapping frames of dimension 60 × 4, repre- senting 0.6 s of IMU data, and a time window was applied to each frame to determine Y at time t by considering the frame representing the 0.6 s of data from t −0.3 s to t +0.3 s. The same parameters used by Gholami et al. [5] were used in this study. Specifically, the first two Conv1D layers had 50 filters followed by a maximum pooling layer of 2:1 subsampling. The following two Conv1D layers had 100 filters, and the outputs were flat- tened and fed into a dense layer of 100 neurons before being consolidated into Y with 3 neurons. A kernel size of 3 with a stride value of 1 and a

first two Conv1D layers had 50 filters followed by a maximum pooling layer of 2:1 subsampling. The following two Conv1D layers had 100 filters, and the outputs were flat- tened and fed into a dense layer of 100 neurons before being consolidated into Y with 3 neurons. A kernel size of 3 with a stride value of 1 and a rectified linear unit, abbreviated as ReLU [44], was used for activation in all layers except the regression output. For opti- mization, Adam [45] was used with a learning rate of 0.001 and batch size of 512 at 50 epochs to achieve rather a stable convergence. For the initialization of neuron weights, the Glorot normal initializer [46] was used (also called the Xavier normal initializer). The Py- thon 3.7 with the Tensorflow and Keras packages were used for implementation. Different arrangements of regressors for X and targets for Y were applied in our study. Since accelerometer readings were likely to deform to different extents on different running surfaces, the gyroscope was also used as a regressor because of the readings being Figure 1.The sensors' placement on the participant.

Sensors2021,21, 4633 4 of 13 2.3. Experiment Protocol and Data Collection All participants were required to complete the Physical Activity Readiness Ques- tionnaire (PAR-Q, Canadian Society for Exercise Physiology, 2002., 30 September 2020) and a medical history questionnaire before the study began. Those with a “Yes” to one or more questions in the PAR-Q, or with any obvious anatomical abnormalities, were excluded. Anthropometric data, including height, weight, and age, were measured and collected. All running trials were conducted on a treadmill in the laboratory. Participants were advised to wear their usual running attire with self-selected running shoes and that they should not participate in any physical activity on the day prior to testing. The IMeasureU and myoMotion sensors were equipped, calibrated, and synchronized. After synchronization, participants warmed up at 1.5 m/s on the treadmill for 30 s. Every 3 min, the speed of the treadmill was increased, rst to 2.0 m/s, then to 2.5 m/s, to 3.0 m/s, and nally to 3.5 m/s [5,43]. The participants were then asked to run in an indoor squash court for 3 min at their preferred speed (average speed: 2.44 0.34 m/s), which was recorded using a Brower Timing System (Brower Timing Systems, Draper, UT, USA). The data obtained from the two sets of IMUs were compared. 2.4. The Deep Learning Regression Model The CNN deep learning model used by Gholami et al. [5] was used to compare the results. Gholami et al. [5] used four layers of one-dimensional convolutional layers (Conv1D) to model the target variable array Y containing the hip, knee, and ankle exion values. A four-column array X was constructed with the rst three columns corresponding to the orthogonal accelerations (ax, ay, and az) and the last column corresponding to the magnitude of the total accelerations, axyz. The rows of X correspond to the IMU data collected at 100 Hz. X was separated into overlapping frames of dimension 60 4, representing 0.6 s of IMU data, and a time window was applied to each frame to determine Y at time t by considering the frame representing the 0.6 s of data

magnitude of the total accelerations, axyz. The rows of X correspond to the IMU data collected at 100 Hz. X was separated into overlapping frames of dimension 60 4, representing 0.6 s of IMU data, and a time window was applied to each frame to determine Y at time t by considering the frame representing the 0.6 s of data from t 0.3 s to t +0.3 s. The same parameters used by Gholami et al. [5] were used in this study. Speci cally, the rst two Conv1D layers had 50 lters followed by a maximum pooling layer of 2:1 subsampling. The following two Conv1D layers had 100 lters, and the outputs were attened and fed into a dense layer of 100 neurons before being consolidated into Y with 3 neurons.A kernel size of 3 with a stride value of 1 and a recti ed linear unit, abbreviated as ReLU [44], was used for activation in all layers except the regression output. For optimization, Adam [45] was used with a learning rate of 0.001 and batch size of 512 at 50 epochs to achieve rather a stable convergence. For the initialization of neuron weights, the Glorot normal initializer [46] was used (also called the Xavier normal initializer). The Python 3.7 with the Tensor ow and Keras packages were used for implementation. Different arrangements of regressors for X and targets for Y were applied in our study. Since accelerometer readings were likely to deform to different extents on different running surfaces, the gyroscope was also used as a regressor because of the readings being less subject to such deformations. For the accelerometer, the y-component, which was more or less aligned with the tibial direction, was used as X. For the gyroscope, the resultant of the x-component and the z-component was used, which more or less re ected the angular motions on the sagittal plane. This selection of regressors was to minimize variations in sensor orientation during the experiments. While Gholami et al. [5] used a combined loss function for the hip, knee, and ankle and used a set of weights to

This study compares accelerometer and gyroscope data for predicting running kinematics.