New Considerations for Collecting Biomechanical Data Using Wearable Sensors: How Does Inclination Influence the Number of Runs Needed to Determine a Stable Running Gait Pattern?

As inertial measurement units (IMUs) are used to capture gait data in real-world environments, guidelines are required in order to determine a `typical' or `stable' gait pattern across multiple days of data collection. Since uphill and downhill running can greatly a ect the biomechanics of running gait, this study sought to determine the number of runs needed to establish a stable running pattern during level, downhill, and uphill conditions for both univariate and multivariate analyses of running biomechanical data collected using a single wearable IMU device. Pelvic drop, ground contact time, braking, vertical oscillation, pelvic rotation, and cadence, were recorded from thirty- ve recreational runners running in three elevation conditions: level, downhill, and uphill. Univariate and multivariate normal distributions were estimated from di ering numbers of runs and stability was de ned when the addition of a new run resulted in less than a 5% change in the 2.5 and 97.5 quantiles of the 95% probability density function for each individual runner. This stability point was determined separately for each runner and each IMU variable (univariate and multivariate). The results showed that 2–4 runs were needed to de ne a stable running pattern for univariate, and 4–5 days were necessary for multivariate analysis across all inclination conditions. Pearson's correlation coe cients were calculated to cross-validate di ering elevation conditions and showed excellent correlations (r=0.98 to 1.0) comparing the training and testing data within the

variable (univariate and multivariate). The results showed that 2–4 runs were needed to de ne a stable running pattern for univariate, and 4–5 days were necessary for multivariate analysis across all inclination conditions. Pearson's correlation coe cients were calculated to cross-validate di ering elevation conditions and showed excellent correlations (r=0.98 to 1.0) comparing the training and testing data within the same elevation condition and good to very good correlations (r=0.63–0.88) when comparing training and testing data from di ering elevation conditions. These results suggest that future research involving wearable technology should collect multiple days of data in order to build reliable and accurate representations of an individual's stable gait pattern. Keywords:accelerometer; gait; elevation; inertial measurement unit; running; slope 1. Introduction As the gait biomechanics research community begins to use inertial measurement units (IMUs) to capture gait data in real-world environments [1–5], guidelines for the use of IMUs during continuous monitoring and multi-day data collections of running gait are required [4]. For example, traditional biomechanics research, involving laboratory-based data collections, generally involve well-controlled external factors, such as temperature and running elevation and only collect a limited number of Sensors2019,19, 2516; doi:10.3390 /s19112516 /journal/sensors

Sensors2019,19, 2516 2 of 9 strides and/or trials [1,4,5]. Considering that changing external factors, such as uctuations in ambient temperature [2] and even whether the activity took place on a weekday or weekend [6] can a ect real-world data collections using IMUs, it is important to quantify the number of sessions needed to de ne an individual's `typical' activity pattern. Until recently it has been suggested that “a few days” of data are necessary in order to establish a “relatively stable” gait pattern [7], however, no speci c information or guidelines were provided. In contrast, a recent study by Benson et al. [5] reported that 4–5 days of data are necessary in order to determine a typical gait pattern across multiple days of data collection. However, three signi cant limitations were apparent in this study that need to be addressed. First, these authors [5] only analysed data from 12 recreational runners, making it di cult to generalise their results to larger populations of runners [5]. Second, data were only collected from runs that were limited to 4–5 km in total distance and only 2.5 km of accumulated data from each run were analysed. Third, and most importantly, these authors constrained their analysis to only level sections of outdoor running and did not consider how running uphill or downhill would in uence the ability to establish a stable gait pattern. It has been shown that uphill and downhill running can greatly a ect gait biomechanics [8–10]. For example, Vernillo et al. [10] reported that the increased demands for work as running slope increases are accompanied by an increase in power output at all joints, particularly the hip, along with an overall progressive move from a rearfoot strike towards a mid- to forefoot strike pattern as running incline increases. In contrast, they reported that the rearfoot strike pattern associated with downhill running was related to increased braking forces. Additionally, it has been reported that uphill or downhill running has several fatigue-related intrinsic features compared with level running that could also a ect gait biomechanics, especially over longer distances [8,11]. For example, Kowalski

forefoot strike pattern as running incline increases. In contrast, they reported that the rearfoot strike pattern associated with downhill running was related to increased braking forces. Additionally, it has been reported that uphill or downhill running has several fatigue-related intrinsic features compared with level running that could also a ect gait biomechanics, especially over longer distances [8,11]. For example, Kowalski and Li [11] reported that when compared to level running, braking forces were larger during downhill conditions and propulsive forces were larger during uphill conditions. Additionally, Ahamed et al. [3] reported that braking force was the most important variable for discriminating between uphill and downhill running for a group of recreational runners. Therefore, considering that uphill and downhill running demand signi cant alterations in gait biomechanics, and considering that uphill and downhill running are commonly experienced by runners in real-world conditions, it remains unknown how many individual runs are required to establish a stable gait pattern during uphill and downhill running compared to that previously reported for level running. The purpose of this study was to determine the number of runs needed to establish a stable running pattern during level, downhill, uphill conditions as well as for an entire run irrespective of elevation (mixed condition). To assess this hypothesis, both the univariate and multivariate analyses of running biomechanical data were collected using a single IMU device. We hypothesized that 4–5 runs would be needed to determine a stable gait patterns for each of the three elevation conditions but that more runs would be necessary for the mixed condition. 2. Materials and Methods 2.1. Participants Based on an a priori power analysis ( =0.05, =0.20), 35 recreational runners (25 females:age=47.6 11.1 years ; height=165.1 5.2 cm; mass=64.7 8.2 kg; and 10 males: age=54.8 9.6 years ; height=174.9 5.8 cm; mass=80.5 7.8 kg) volunteered to participate in this study. The runners were free of any neuromuscular diseases or musculoskeletal injuries and were registered for a marathon training program managed by a local running group. This protocol was approved by the University of Calgary Conjoint Health Research Ethics Board (REB16-2035), and all the

males: age=54.8 9.6 years ; height=174.9 5.8 cm; mass=80.5 7.8 kg) volunteered to participate in this study. The runners were free of any neuromuscular diseases or musculoskeletal injuries and were registered for a marathon training program managed by a local running group. This protocol was approved by the University of Calgary Conjoint Health Research Ethics Board (REB16-2035), and all the runners provided their written informed consent.

Sensors2019,19, 2516 3 of 9 2.2. Instrumentation Biomechanical gait variables from each runner were recorded using a commercially available wearable IMU (Lumo Run ® ; Lumo Bodytech Inc., Mountain View, CA, USA; sampling rate of 100 Hz and recording data every ve strides), consisting of a three-dimensional accelerometer, magnetometer, and gyroscope. According to the manufacturer's instructions, the IMU was attached to the back of the shorts or to a running belt near the individual's center of mass. During each run, the IMU performed on-board computation of six biomechanical variables: pelvic drop (deg), vertical oscillation (cm), ground contact time (ms), braking (m/s), pelvic rotation (deg) and cadence (steps/min). A GPS watch (Garmin v½voactive ® HR; Garmin International Inc., Olathe, KS, USA) was attached to each runner's preferred wrist and recorded data (1 Hz) on the runner's speed (m/s), distance (km), and global positioning data, including latitude, longitude and altitude (Supplementary S2). For each run, the runner's clothes, shoes, device placement, pace and running route were not controlled. A MATLAB-based (MATLAB ® 2017a) program was developed to synchronize the IMU and GPS data for further analysis. 2.3. Data Collection Runners wore both the IMU and the GPS watch during all training runs over the course of their marathon training program. Seven pain-free runs, of at least 12 km in length, were performed by each runner and used in the analysis. For each runner, no more than 1 run was recorded on a single day with an average of 12.9 days ( 6.54) between the collective runs. Data from each individual run were considered based on the following four conditions; (i) mixed condition: (the entire run and irrespective of elevation), (ii) level: ( 2%), (iii) uphill: (+3% to+15%), and (iv) downhill: ( 3% to 15%) running conditions. To minimize the potential e ects of fatigue, only data from the rst 10 km of each run were used with data from the rst kilometer discarded and considered as a warmup period. Each run was segmented into elevation sections of at least 100 m in length. A 10 s moving average was used to smooth

to 15%) running conditions. To minimize the potential e ects of fatigue, only data from the rst 10 km of each run were used with data from the rst kilometer discarded and considered as a warmup period. Each run was segmented into elevation sections of at least 100 m in length. A 10 s moving average was used to smooth elevation and speed data from the GPS watch and sections where running speed was less than 1.8 m/s [12] were removed so that only data where subjects were actively running was used in the analysis. The average elevation gain or loss was computed for continuous non-overlapping windows of 100 m length, which were then segmented into each terrain type. A summary of these data is provided in Table. Table 1. Means and standard deviations of the characteristics for each of the four running conditions. Elevation Condition Total Distance (km) Average Inclination (%) Speed (m /s) Mean SD Mean SD Mean SD Level 4.2 2.1 1.88 1.44 2.41 0.26 Uphill 3.3 1.1 3.24 1.61 2.43 0.25 Downhill 3.5 1.3 3.81 2.52 2.44 0.22 Mixed 8.3 3.2 0.17 2.51 2.45 0.25 2.4. Data Analysis A detailed description of the data analysis process is described in detail elsewhere [5]. In brief, to determine the number of runs needed to establish a stable running pattern, training datasets consisting of varying numbers of runs were created, and the runs that did not form part of a training dataset comprised the corresponding testing datasets. A leave-one-out cross-validation approach was used to determine the level of similarity between all training sets from the thirty- ve runners for each of the four conditions and a test run performed in the same condition. Thus, 7-N elevation-speci c testing datasets were obtained for each unique training dataset, where N is the number of runs in the training dataset (Table). For each unique training dataset, the 95% probability density function

Sensors2019,19, 2516 4 of 9 of the univariate or multivariate normal distribution was determined using the multivariate normal probability density function (MVNPDF) (Supplementary S1) function in MATLAB [13,14], and these analyses were performed for each of the IMU variables individually (univariate analyses) and for all of the IMU variables together (multivariate analyses) (Supplementary S3). Table 2.Unique sets of runs based on number of runs (train-test pairs) included in each set. Number of Runs Per Set Number of Unique Sets Sets Number of Training-Testing Dataset Pairs 1 7 {Run 1}, {Run 2}, : : : 49 2 21 {Runs 1 +2}, {Runs 1+3},: : : 126 3 35 {Runs 1 +2+3}, {Runs 1+2+4},: : : 175 4 35 {Runs 1 +2+3+4}, {Runs 1+2+3+5},: : : 140 5 21 {Runs 1 +2+3+4+5}, {Runs 1+2+3+4+6},: : : 63 6 7 {Runs 1+2+3+4+5+6}, {Runs 1+2+3+4+5+7}, 14 7 1 {Runs 1 +2+3+4+5+6+7} 0 A threshold of the resulting 95% probability density function, epsilon, was determined such that 95% of the training data had a probability greater than or equal to epsilon. The data points in the testing dataset with a probability greater than or equal to epsilon were considered similar to the training dataset, and the percentage of similar data points in the testing dataset was recorded. Univariate and multivariate normal distributions were estimated from di ering numbers of runs and stability was de ned when the addition of a new run resulted in less than a 5% change in the 2.5 and 97.5 quantiles of the estimated density for each individual runner. This stability point was determined separately for each runner and each IMU variable (univariate and multivariate). The overall stability point for each IMU variable was determined as the maximum number of runs needed to reach stability among the thirty- ve runners for each of the four elevation conditions. In order to further cross-validate our approach, a crossover condition analysis was done by randomly selecting five runs for each individual as training runs, with the remaining two runs designated as test runs. For each condition, a model was trained using data from the

needed to reach stability among the thirty- ve runners for each of the four elevation conditions. In order to further cross-validate our approach, a crossover condition analysis was done by randomly selecting five runs for each individual as training runs, with the remaining two runs designated as test runs. For each condition, a model was trained using data from the training runs only, and data in each condition from the testing runs was used as separate test sets (e.g., train on uphill from five training runs, separate tests with uphill, downhill, level, and all from two testing runs). The number of data points outside the 95% probability density function of the trained model was the outcome variable. This procedure was repeated so that a model was trained for each of the four conditions, and for all variables. The association between the number of data points outside the 95% probability density function for each train/test combination in the crossover condition analysis and the number of data points outside the 95% probability density function at the five-run stability point in the original analysis was determined using Pearson's correlation coefficient. Based on previous studies, we defined r<0.60 as poor, 0.61–0.80 as good, 0.81–0.95 as very good, and>0.95 as excellent [15,16]. All statistical tests were conducted using Minitab ® statistical software (Minitab Inc., State College, PA, USA). 3. Results The average and maximum number of runs needed to de ne stable running patterns for each elevation condition, and for the data from the entire run, are shown in Table. Overall, the results showed that 2–4 days were needed to de ne a stable running pattern for univariate and 4–5 days were necessary for multivariate analysis. Cadence required the fewest number of days to reach stability especially during downhill running (two days) compared to level and uphill running (three days) and also compared to the other univariate measures (3–4 days). The multivariate analysis required 4–5 days to reach stability irrespective of elevation condition.

stability especially during downhill running (two days) compared to level and uphill running (three days) and also compared to the other univariate measures (3–4 days). The multivariate analysis required 4–5 days to reach stability irrespective of elevation condition.

Sensors2019,19, 2516 5 of 9 Table 3. Number of days (mean, standard deviation (SD) and maximum) needed to de ne a stable running pattern (univariate and multivariate). Units Variable /Condition Cadence Bounce Braking Pelvic Drop Pelvic Rotation Ground Contact Time Multivariate Mean (SD) Level 1.71 (0.67) 2.09 (0.61) 2.46 (0.61) 2.60 (0.81) 2.26 (0.71) 2.34 (0.76) 3.29 (0.79) Uphill 1.74 (0.56) 2.10 (0.66) 2.61 (0.85) 2.49 (0.82) 2.11 (0.68) 2.17 (0.75) 3.14 (0.65) Downhill 1.37 (0.49) 2.11 (0.63) 2.46 (0.78) 2.40 (0.74) 1.97 (0.66) 2.29 (0.57) 3.20 (0.53) Mixed 1.69 (0.63) 2.17 (0.82) 2.26 (0.56) 2.63 (0.73) 2.27 (0.71) 2.23 (0.59) 3.21 (0.72) Max Level 3 3 3 4 4 4 5 Uphill 3 3 4 4 3 3 4 Downhill 2 3 4 4 3 3 4 Mixed 3 4 3 4 4 3 5 The cross-validation based on the Pearson's correlation coe cient and comparison of the average number of data points that fell outside of the 95% probability density function showed excellent to perfect correlations (r=0.98 to 1.0) comparing the training and testing data within the same elevation condition (Table). However, if the training and testing data were from di erent elevation conditions, the correlations were considered to be between good and very good (r=0.63 to 0.88) for level, uphill, and downhill conditions and also good to very good for the mixed condition (r=0.68 to 0.92). Table 4. Correlation matrices of similar data points between training and testing data sets. Grey cells indicate resultant r-values when training and testing data sets were from the same running conditions Variables Train Test Level UphillDownhill Mixed Cadence Level 0.98 0.75 0.76 0.88 Uphill 0.77 1.0 0.67 0.86 Downhill 0.67 0.66 1.0 0.82 Mixed 0.82 0.85 0.85 0.99 Vertical Oscillation Level 0.99 0.76 0.69 0.92 Uphill 0.74 0.99 0.71 0.91 Downhill 0.63 0.73 1.0 0.85 Mixed 0.81 0.82 0.68 1.0 Braking Level 1.0 0.66 0.73 0.84 Uphill 0.78 1.0 0.76 0.86 Downhill 0.71 0.74 0.99 0.84 Mixed 0.85 0.84 0.81 1.0 Pelvic Drop Level 0.99 0.68 0.72 0.81 Uphill 0.75 0.98 0.68 0.83 Downhill 0.76 0.75 1.0 0.84 Mixed 0.83 0.81

0.99 0.76 0.69 0.92 Uphill 0.74 0.99 0.71 0.91 Downhill 0.63 0.73 1.0 0.85 Mixed 0.81 0.82 0.68 1.0 Braking Level 1.0 0.66 0.73 0.84 Uphill 0.78 1.0 0.76 0.86 Downhill 0.71 0.74 0.99 0.84 Mixed 0.85 0.84 0.81 1.0 Pelvic Drop Level 0.99 0.68 0.72 0.81 Uphill 0.75 0.98 0.68 0.83 Downhill 0.76 0.75 1.0 0.84 Mixed 0.83 0.81 0.65 1.0 Pelvic Rotation Level 1.0 0.76 0.69 0.84 Uphill 0.72 0.99 0.67 0.87 Downhill 0.73 0.72 0.99 0.86 Mixed 0.81 0.85 0.86 1.0 Ground Contact Time Level 0.99 0.72 0.66 0.70 Uphill 0.74 1.0 0.68 0.89 Downhill 0.71 0.71 1.0 0.82 Mixed 0.84 0.83 0.88 1.0 Multivariate Level 0.98 0.63 0.74 0.73 Uphill 0.62 1.0 0.76 0.66 Downhill 0.70 0.73 0.99 0.68 Mixed 0.67 0.71 0.69 1.0 4. Discussion The purpose of the current study was to determine the number of runs needed to establish a stable or typical running pattern during real-world running conditions of level, uphill, and downhill running grade based on both univariate and multivariate analyses. In partial support of our hypotheses, the rst

The study investigates the impact of running inclination on gait stability using wearable sensors.