Estimating Lower Extremity Running Gait Kinematics with a Single Accelerometer: A Deep Learning Approach

Abnormal running kinematics are associated with an increased incidence of lower extremity injuries among runners. Accurate and unobtrusive running kinematic measurement plays an important role in the detection of gait abnormalities and the prevention of injuries among runners. Inertial-based methods have been proposed to address this need. However, previous methods require cumbersome sensor setup or participant-speci c calibration. This study aims to validate a shoe-mounted accelerometer for sagittal plane lower extremity angle measurement during running based on a deep learning approach. A convolutional neural network (CNN) architecture was selected as the regression model to generalize in inter-participant scenarios and to minimize poorly estimated joints. Motion and accelerometer data were recorded from ten participants while running on a treadmill at ve di erent speeds. The reference joint angles were measured by an optical motion capture system. The CNN model predictions deviated from the reference angles with a root mean squared error (RMSE) of less than 3.5 and 6.5 in intra- and inter-participant scenarios, respectively. Moreover, we provide an estimation of six important gait events with a mean absolute error of less than 2.5 and 6.5 in intra- and inter-participants scenarios, respectively. This study highlights an appealing minimal sensor setup approach for gait analysis purposes. Keywords: accelerometer; inertial sensors; wearable sensors; gait monitoring; kinematic; running; convolutional neural networks 1. Introduction The kinematics of the lower extremity are an important area of research in human running gait analysis. Abnormal running kinematics are associated with an increased incidence of lower extremity injuries among runners [1,2]. While injuries may occur due to deviations in any plane, two of the most common kinematic patterns

sensors; gait monitoring; kinematic; running; convolutional neural networks 1. Introduction The kinematics of the lower extremity are an important area of research in human running gait analysis. Abnormal running kinematics are associated with an increased incidence of lower extremity injuries among runners [1,2]. While injuries may occur due to deviations in any plane, two of the most common kinematic patterns associated with injured runners occur in the sagittal plane (greater knee extension and ankle dorsi exion at initial contact [2]). Correcting abnormal kinematics has been suggested to reduce the risk of injury [3,4]. Runners' gait parameters can also change during a prolonged run due to fatigue [5–7], which may increase injury risk as the runner deviates from their normal gait pattern [8,9]. Reducing injury risk factors via altering running biomechanics has been demonstrated to be feasible in a lab environment [4,10], but this requires sophisticated equipment and software. Providing this feedback in an ecologically valid (in- eld) context requires the ability to both measure and give feedback using wearable devices. This can be done currently only by using simple metrics such as cadence [11] or peak tibial acceleration [12], but the measurement of joint kinematics Sensors2020,20, 2939; doi:10.3390 /s20102939 /journal/sensors

Sensors2020,20, 2939 2 of 14 is much more complex. Developing an accurate, unobtrusive gait monitoring wearable system to measure running kinematics is still subject to research. In contrast to optical motion-capture-based gait analysis, wearable sensors enable continuous and unobtrusive gait monitoring during in- eld activity. Inertial measurement units (IMUs) have been widely employed as a portable system for the estimation of human gait kinematics and kinetics [13,14]. Consisting of an accelerometer, gyroscope, and magnetometer, IMUs measure the orientation of body segments by sensor data fusion. 1.1. IMU-Based Gait Kinematic Estimation Inertial-based sensors have been employed to measure human kinematics based on three main approaches: (1) orientation-based methods that entail calculating the relative orientation of two adjacent segments based on the orientation of IMUs mounted on the distal and proximal segments; (2) model-based methods that take advantage of kinematic constraints; and (3) data-driven methods that employ supervised learning models to estimate reference joint angles. The orientation-based method has traditionally been used to compute human kinematics [13,15] but has some limitations, including errors in sensor-to-segment alignment. The anatomical orientation of the bone is di erent from the local reference frame of the IMUs, requiring a calibration step to calculate the relative orientation of the IMU and the anatomical reference frame. However, the precision of the calibration also relies on performing prede ned movements correctly, which may not be reliable, especially in injured populations [15,16]. Previous studies have used motion capture or RGB cameras to calculate the relative orientation of IMUs and the anatomical position before testing [17–20]. Although this method obtains lower errors, it requires prior calibration with cameras for each new session or participant. Another limitation of the orientation-based method is the drift in sensor signal that is due to signal integration operation, which is usually an essential step. Sensor data fusion using the Kalman lter is the main approach that has been proposed to overcome the drift problem [21]. However, the signal drift in noisy environments with ferromagnetic disturbances and during prolonged data recording is still problematic [13]. Kinematic or musculoskeletal model-based approaches consider the kinematic constraints of the human

integration operation, which is usually an essential step. Sensor data fusion using the Kalman lter is the main approach that has been proposed to overcome the drift problem [21]. However, the signal drift in noisy environments with ferromagnetic disturbances and during prolonged data recording is still problematic [13]. Kinematic or musculoskeletal model-based approaches consider the kinematic constraints of the human body to overcome the limitations of sensors such as drift and sensor-to-segment alignment. Optimization is driven to reduce the error between the orientation of the IMUs and body segments by taking into account the anatomical constraints of the human body [22]. Dorschky et al. [23] proposed estimating lower extremity joint angles using raw accelerometer and gyroscope data by solving optimization problems based on the musculoskeletal dynamics model introduced by van denBogert et al. [24]. Joint constraints have also been proposed to overcome sensor-to-segment calibration [16]. The third method for lower extremity joint angle monitoring is a data-driven approach that relies on machine learning algorithms [25–27]. This approach feeds the raw signal [26] or orientation calculated by ltering methods into a machine learning model [27]. The machine learning model takes care of the sensor-to-segment calibration and calculates the joint angles. The main limitation of the data-driven method is the dependency on the supplied dataset and generalizability to other individuals or populations. Wouda et al. [27] used three IMUs to measure ground reaction forces and knee joint angles in running based on a neural network model. They demonstrated the potential of data-driven methods to estimate both kinematic and kinetic parameters during highly dynamic movement. Previous studies have reported root mean squared error (RMSE) values for sagittal plane lower extremity joint angles in walking ranging from 1.5 to 11 [15,28,29], while in running a range of 3.4 –13 for hip, knee, and ankle angle estimation has been reported [17,27]. When comparing the ndings of previous studies, it is important to mention that many studies have included prior information for the model, such as a participant-speci c optimized model [23], sensor-to-segment calibration [17,19,20,30], or sensor initial orientation on the body segment [20,31].

a range of 3.4 –13 for hip, knee, and ankle angle estimation has been reported [17,27]. When comparing the ndings of previous studies, it is important to mention that many studies have included prior information for the model, such as a participant-speci c optimized model [23], sensor-to-segment calibration [17,19,20,30], or sensor initial orientation on the body segment [20,31].

Sensors2020,20, 2939 3 of 14 1.2. Reducing Number and Degrees of Freedom of IMUs: Mounting an IMU on each segment for lower extremity kinematic estimation requires seven IMUs to provide full lower extremity joint angle measurement. This approach is cumbersome and obtrusive. Furthermore, it is unrealistic to expect a user at the consumer level to attach multiple IMUs due to the increased cost, reduced practicality, and probability of xation error. Hence, there is a trend to reduce the number of required sensors for joint angle estimation. There are two approaches proposed so far to reduce the number of sensors: (1) model-based and (2) data-driven. A model-based approach was proposed by Hu et al. [32] using four IMUs only to estimate the hip, knee, and ankle angles based on a serial chain model and solving the inverse kinematics. Bonnet et al. [31] used a single IMU on the shank to measure the hip and knee angles during several rehabilitation exercises by considering the mechanical constraints between proximal and distal segments. These methods demonstrated a greater error compared to a comprehensive sensor setup or were introduced for speci c tasks. Data-driven methods have also been proposed for human full-body motion monitoring [33,34] and lower extremity walking analysis [35] with a reduced number of sensors. Recently, Lim et al. [35] proposed using a single IMU on the pelvis close to the center of mass to measure the lower limb kinematics and kinetics in walking. However, the application of a reduced number of inertial sensors to measure gait kinematics during a highly dynamic motion such as running has not yet been investigated. 1.3. Other Considerations One issue with the use of IMUs for joint orientation measurement is the sensitivity of the magnetometer to ferromagnetic disturbances. For this reason, there is a trend towards magnetometer- free inertial-based systems for joint angle monitoring. Power consumption is also a signi cant concern during activities that occur over a prolonged period. Long distance running can have a duration of up to two hours or more. When comparing the power consumption of accelerometers and gyroscopes, the former has the

disturbances. For this reason, there is a trend towards magnetometer- free inertial-based systems for joint angle monitoring. Power consumption is also a signi cant concern during activities that occur over a prolonged period. Long distance running can have a duration of up to two hours or more. When comparing the power consumption of accelerometers and gyroscopes, the former has the advantage of a lower power consumption [36]. Therefore, the use of an IMU consisting of an accelerometer alone has the advantage of avoiding ferromagnetic disturbances and lasting throughout longer-duration activities. The anatomical location of the IMU impacts the practicality of its use as well as the quality of the data. Rigid xation of IMUs on the shank or thigh is less practical and more di cult than using a shoe-mounted system. The xation that can be applied using a shoe- or lace-mounted IMU is likely to lead to more reliable data. The ability to measure running kinematics outside of the lab and in real-world environments has the potential not only to allow researchers to study the impact of di erent terrain and fatigue states on running biomechanics, but also to prevent injuries when paired with real-time biofeedback [37]. In order to achieve this goal, accurate and unobtrusive monitoring of kinematics needs to be demonstrated in a controlled setting. In this study, we aimed to investigate the performance of a single shoe-mounted accelerometer to monitor lower extremity running kinematics in the sagittal plane. A secondary aim was to reduce the degrees of freedom of the IMU. A data-driven approach is presented based on convolutional neural networks. The performance of the method in inter- and intra-participant scenarios has been evaluated. 2. Materials and Methods 2.1. Experiment Setup Twenty- ve re ective markers were a xed to each participant prior to testing, and a static calibration trial was initially collected to form a musculoskeletal model based on Napier et al. (Figure) [ 38] using a 6-camera motion analysis system (Vicon, Oxford, UK). Ten static/calibration markers (anterior superior iliac spines, greater trochanters, left medial/lateral femoral condyles, left medial/lateral malleoli, and left rst and fth

were a xed to each participant prior to testing, and a static calibration trial was initially collected to form a musculoskeletal model based on Napier et al. (Figure) [ 38] using a 6-camera motion analysis system (Vicon, Oxford, UK). Ten static/calibration markers (anterior superior iliac spines, greater trochanters, left medial/lateral femoral condyles, left medial/lateral malleoli, and left rst and fth metatarsal heads) were removed following the static

Sensors2020,20, 2939 4 of 14 trial, and the 15 remaining markers (posterior superior iliac spines, iliac crests, clusters of 4 on the thigh and shank, and a triad on the heel) were by de nition tracking and calibration markers, as they were on for both static and dynamic trials. The ankle joint center was identi ed as the midpoint between the 2 ankle malleoli markers; the knee joint center was identi ed as the midpoint between the 2 femoral condyle markers, and the hip joint center was identi ed using the method of Bell et al. [39].Sensors 2020, 20, x FOR PEER REVIEW 4 of 14 between the 2 ankle malleoli markers; the knee joint center was identified as the midpoint between the 2 femoral condyle markers, and the hip joint center was identified using the method of Bell et al. [39]. (A) ( B) Figure 1. (A) Experimental setup including six motion capture cameras and a split-belt treadmill. The schematic of angles estimated using the raw signal of a foot-mounted accelerometer. (B) Reflective marker positions on the lower extremity. The acceleration of the foot was measured by an Xsens inertial measurement unit (MTw Awinda, Xsens, Enschede, The Netherlands) mounted on the shoe, as shown in Figure 1. The Xsens unit has an accelerometer, gyroscope, and magnetometer; however, in this study only the raw data of the accelerometer were used for data processing. Both the motion capture and accelerometer data were recorded at a 100 Hz sampling rate. The accelerometer and motion capture data were synchronized using an analog signal from each of the Vicon and Xsens units. 2.2. Data Collection Ten healthy male Caucasian participants (age 27 ± 4 years, height 177 ± 7 cm, and weight 72 ± 7 kg) participated in this study. The number of participants (sample size) was conducted in G*Power 3.1.9.3 to detect a strong (R 2 > 0.80) association between our method and the gold standard optical motion capture method for computing joint kinematics. To obtain 80% power to detect significant (p < 0.05) associations, we determined that 10 participants were required. Sex

7 kg) participated in this study. The number of participants (sample size) was conducted in G*Power 3.1.9.3 to detect a strong (R 2 > 0.80) association between our method and the gold standard optical motion capture method for computing joint kinematics. To obtain 80% power to detect significant (p < 0.05) associations, we determined that 10 participants were required. Sex was not expected to influence our results [40]. The study protocol was approved by the Office of Research Ethics at Simon Fraser University, and all participants provided informed consent. The data recording protocol consisted of 15 trials of running at five different speeds—8, 9, 10, 11, and 12 km/h—with three trials of 60 s at each speed. The participants were given time to warm up and familiarize themselves with the treadmill before data recording started from the slowest speed. The participants were given a short break after each trial. 2.3. Data Preprocessing The raw accelerometer data from all three axes were recorded by Xsens software and filtered using the SciPy Python library [41]. The accelerometer data were not normalized or standardized. A total of 6% of the dataset was excluded due to the asynchronized accelerometer and motion signals. The motion capture data were considered the gold standard reference for the kinematic data for this study. The marker trajectories were imported to Visual 3D software (C-Motion, Inc., Germantown, MD, USA) and the joint angles were computed and filtered with Visual 3D. The motion and accelerometer data were filtered by a fourth-order Butterworth low-pass filter with a cut-off frequency of 6 Hz following the recommendation of previous studies [42]. A sample of raw accelerometer data is shown in Figure 2. Figure 1. (A) Experimental setup including six motion capture cameras and a split-belt treadmill. The schematic of angles estimated using the raw signal of a foot-mounted accelerometer. (B) Re ective marker positions on the lower extremity. The acceleration of the foot was measured by an Xsens inertial measurement unit (MTw Awinda, Xsens, Enschede, The Netherlands) mounted on the shoe, as shown in Figure. The Xsens unit has an accelerometer, gyroscope, and magnetometer;

split-belt treadmill. The schematic of angles estimated using the raw signal of a foot-mounted accelerometer. (B) Re ective marker positions on the lower extremity. The acceleration of the foot was measured by an Xsens inertial measurement unit (MTw Awinda, Xsens, Enschede, The Netherlands) mounted on the shoe, as shown in Figure. The Xsens unit has an accelerometer, gyroscope, and magnetometer; however, in this study only the raw data of the accelerometer were used for data processing. Both the motion capture and accelerometer data were recorded at a 100 Hz sampling rate. The accelerometer and motion capture data were synchronized using an analog signal from each of the Vicon and Xsens units. 2.2. Data Collection Ten healthy male Caucasian participants (age 27 4 years, height 177 7 cm, and weight 72 7 kg) participated in this study. The number of participants (sample size) was conducted in G*Power 3.1.9.3 to detect a strong (R 2 >0.80) association between our method and the gold standard optical motion capture method for computing joint kinematics. To obtain 80% power to detect signi cant (p<0.05) associations, we determined that 10 participants were required. Sex was not expected to in uence our results [40]. The study protocol was approved by the O ce of Research Ethics at Simon Fraser University, and all participants provided informed consent. The data recording protocol consisted of 15 trials of running at ve di erent speeds—8, 9, 10, 11, and 12 km/h—with three trials of 60 s at each speed. The participants were given time to warm up and familiarize themselves with the treadmill before data recording started from the slowest speed. The participants were given a short break after each trial. 2.3. Data Preprocessing The raw accelerometer data from all three axes were recorded by Xsens software and ltered using the SciPy Python library [41]. The accelerometer data were not normalized or standardized. A total of 6% of the dataset was excluded due to the asynchronized accelerometer and motion signals. The motion capture data were considered the gold standard reference for the kinematic data for this study. The marker trajectories were

axes were recorded by Xsens software and ltered using the SciPy Python library [41]. The accelerometer data were not normalized or standardized. A total of 6% of the dataset was excluded due to the asynchronized accelerometer and motion signals. The motion capture data were considered the gold standard reference for the kinematic data for this study. The marker trajectories were imported to Visual 3D software (C-Motion, Inc., Germantown, MD, USA) and the joint angles were computed and ltered with Visual 3D. The motion and accelerometer data

The study investigates a deep learning method for measuring running kinematics using a single accelerometer.