How Precisely Can Easily Accessible Variables Predict Achilles and Patellar Tendon Forces during Running?

ellar and Achilles tendinopathy commonly affect runners. Developing algorithms to predict cumulative force in these structures may help prevent these injuries. Importantly, such algorithms should be fueled with data that are easily accessible while completing a running session outside a biomechanical laboratory. Therefore, the main objective of this study was to investigate whether algorithms can be developed for predicting patellar and Achilles tendon force and impulse during running using measures that can be easily collected by runners using commercially available devices. A secondary objective was to evaluate the predictive performance of the algorithms against the commonly used running distance. Trials of 24 recreational runners were collected with an Xsens suit and a Garmin Forerunner 735XT at three different intended running speeds. Data were analyzed using a mixed-effects multiple regression model, which was

be easily collected by runners using commercially available devices. A secondary objective was to evaluate the predictive performance of the algorithms against the commonly used running distance. Trials of 24 recreational runners were collected with an Xsens suit and a Garmin Forerunner 735XT at three different intended running speeds. Data were analyzed using a mixed-effects multiple regression model, which was used to model the association between the estimated forces in anatomical structures and the training load variables during the xed running speeds. This provides twelve algorithms for predicting patellar or Achilles tendon peak force and impulse per stride. The algorithms developed in the current study were always superior to the running distance algorithm. Keywords: Garmin; wearables; Achilles tendon; patellar tendon; algorithm; injuries; sports medicine 1. Introduction Exercise should be taken seriously since it is believed to have profound health ben- e ts [1]. One type of exercise activity is running, which, on a global scale, has gained popularity in the past decades. Running is preferred by many, owing to its accessibility and bene cial effects on various health-related outcomes, such as tness level and health [2]. In contrast to its bene ts, running can also lead to injuries in the musculoskeletal system [3]. Patellar and Achilles tendinopathy account for more than 10% of all running- related injuries [4]. These conditions and running-related injury, in general, are major obstacles to exercise activity [5], so prevention of patellar and Achilles tendon injuries are important. In-depth knowledge about forces applied to the involved anatomical structures is needed because an injury occurs when the cumulative tendon load exceeds the structure's capacity to withstand the load [6–8]. Cumulative tendon load is considered a superior metric for the prediction of injury compared to running distance [9], which has been widely used in the previous literature [10]. Therefore, more sophisticated measures of load are warranted. An advanced method to quantify load is to estimate tendon force using computational models of the musculoskeletal system [11]. Unfortunately, this method seems practically Sensors2021,21, 7418.

which has been widely used in the previous literature [10]. Therefore, more sophisticated measures of load are warranted. An advanced method to quantify load is to estimate tendon force using computational models of the musculoskeletal system [11]. Unfortunately, this method seems practically Sensors2021,21, 7418.

Sensors2021,21, 7418 2 of 12 and logistically impossible in a real-life setting due to high computational complexity and demands of detailed motion data. Therefore, developing a computationally simple algorithm to predict the cumulative force in the Achilles or the patellar tendon is an important step to improving the understanding of the etiology underpinning running injury in these structures. If successful, such algorithms can be used to obtain session- speci c and structure-speci c approximations of tissue loads in large-scale epidemiological studies examining the “too much training load, too soon”-theory [7]. Large-scale studies are needed to assess changes in the tendon force in different groups displaying different recovery patterns [7], running experience [6], previous injuries [6], and pain sensitivity [12], to name a few. In the ideal study, thousands of runners should be included to consider various effect-measure modi ers. However, such a sample size is likely to make it dif cult to obtain full-body kinematic and kinetic data in a time-ef cient manner. To obtain enough personalized data on individual runners, an algorithm predicting cumulative force on an Achilles or patellar tendon should rely on measures that are easy to assess in-situ, rather than advanced measures that can only be assessed in biomechanical laboratories. As an example, running measures (e.g., speed, cadence), which are measurable by commercially available devices, such as smartwatches, may be used when developing algorithms to predict step-speci c forces in the patellar and Achilles tendons. Commercially available devices are widely used by runners, as they can be worn unobtrusively during running [13,14]. These devices have made it possible to obtain indirect measures of training load (such as the number of strides, cadence, ground contact time, and vertical oscillation) in an outdoor environment [15,16]. Using such measures to calculate approximations of forces in the patellar and Achilles tendon force requires, however, that the approximations have acceptable predictability of biomechanically-assessed forces [6,17]. Therefore, the main objective of the present study was to investigate whether an algo- rithm can be developed to estimate patellar and Achilles tendon forces and impulses per stride during running, using measures that can be

to calculate approximations of forces in the patellar and Achilles tendon force requires, however, that the approximations have acceptable predictability of biomechanically-assessed forces [6,17]. Therefore, the main objective of the present study was to investigate whether an algo- rithm can be developed to estimate patellar and Achilles tendon forces and impulses per stride during running, using measures that can be easily collected by Garmin devices with- out expert assistance. A secondary objective was to evaluate the predictive performance of the algorithm while estimating patellar and Achilles tendon peak force and impulse per stride and compare the algorithm with running distance metrics, which is the common way in the literature of estimating cumulative load and serves as a control. 2. Materials and Methods 2.1. Subjects Twenty-four runners (17 males and 7 females) were included in the study with an average age of 26 1.3 years. The runners weighed 82 11 kg; stature 182 7 cm; and had a knee, ankle, and shoe sole height of 49.6 3.1 cm, 7.6 0.9 cm, and3 0.9 cm, respectively (see Supplementary Material, Table S1 for more detailed numbers). All runners were recreationally active for at least 60 min per week. Additionally, all runners had been injury-free for at least six months and completed the setup described in the protocol without any complaints, pain, or discomfort. Before testing, the runners were informed about the purpose of the study, study design, equipment, and signed a declaration of informed written consent. The Regional Ethics Committee of North Jutland waived the approval of the study owing to the study design (an observational cross-sectional study) since observational studies do not require approval from the local ethics committee according to Danish Law. 2.2. Procedures Prior to data collection, each runner was introduced to the Garmin Forerunner 735XT (GFR) to become familiar with the start, stop and save function. Anthropometric data for each runner were collected based on instructions provided by Xsens (Xsens Technologies B.V, Enschede, The Netherlands). Seventeen inertial motion units (IMU) were mounted on the Xsens suit on the following anatomical locations: head, sternum, pelvis, upper legs, lower legs, feet,

introduced to the Garmin Forerunner 735XT (GFR) to become familiar with the start, stop and save function. Anthropometric data for each runner were collected based on instructions provided by Xsens (Xsens Technologies B.V, Enschede, The Netherlands). Seventeen inertial motion units (IMU) were mounted on the Xsens suit on the following anatomical locations: head, sternum, pelvis, upper legs, lower legs, feet, shoulders, upper arms, forearms, and hands using the designated clothing

Sensors2021,21, 7418 3 of 12 items of the Xsens system. The anthropometric data were loaded into the Xsens MVN Studio 4.3 before calibration. Prior to the data collection, the runners performed a 20-min warm-up at a self-selected pace, followed by ve-min rest. During this resting period, a calibration of the inertial motion capture system was performed. Segment orientations were obtained by applying the IMU-to-segment alignment, found using a known upright pose (N-pose) [18]. This N-pose calibration updates the joints and external contacts to limit the position drift [18]. If the calibration was categorized as “good” according to the system, it was accepted and redone otherwise. A visual inspection using the live view of the joint movement was performed to ensure that the recorded data were consistent with the movement of the runner. The calibration was performed outdoor to reduce the amount of magnetic disturbance [18]. Roetenberg, Luinge, and Slycke provide a detailed description of the Xsens system [18]. The runner was equipped with the GFR after the calibration was completed. 2.3. Data Collection Each of the runners performed three running trials of two minutes on a straight outdoor track paved with asphalt at three different speeds (10, 12, and 14 km/h) in randomized order. The speed was controlled by a person riding a bike in front of the runner with a Garmin Fenix 2 GPS watch (Garmin Ltd., Olathe, KS, USA) mounted on the bike, while another rode behind the runner on a Long John bicycle with a computer, a battery and access point for data collection. The Xsens system was activated rst. Then, the runner was instructed to perform a jump and start the GFR upon landing to synchronize the two datasets. 3D kinematic data of the full-body were recorded at 240 Hz with an Xsens MVN link motion capture suit. Running dynamic data were recorded each second (60 Hz) with the GFR paired with a heart rate strap (HRM-Run; Garmin Ltd., Olathe, KS, USA). The GFR variables that were measured during the run and included in the present study were: instantaneous Speed, Vertical Oscillation, Ground Contact Time, Step

recorded at 240 Hz with an Xsens MVN link motion capture suit. Running dynamic data were recorded each second (60 Hz) with the GFR paired with a heart rate strap (HRM-Run; Garmin Ltd., Olathe, KS, USA). The GFR variables that were measured during the run and included in the present study were: instantaneous Speed, Vertical Oscillation, Ground Contact Time, Step length, and Cadence. 2.4. Data Processing Data from GFR were downloaded using Garmin Connect software 7.1.4.0 (Garmin Ltd., Olathe, KS, USA) and exported to R (v. 4.0.5). The Xsens data was captured in a native le format called MVN), which was HD reprocessed in MVN studio to get the best performance for the recorded motion. The data were aligned using the jump as an indicator of the start of the Garmin data. Furthermore, 35 s of data were disregarded from the start of both systems to make sure the runner had reached a constant speed. The kinematic data from Xsens were exported in Biovision Hierarchy (BVH) format and thereafter processed in a computer model (BVH_Xsens template in AnyBody Managed Model Repository version 2.2) of the musculoskeletal system using the AnyBody Modeling System (version 7.2). In the AnyBody Modeling System, the patellar and Achilles tendon forces were estimated for four strides per trial [19–22], using the muscle recruitment ap- proach described in Damsgaard et al. [21]. We predicted ground reaction forces from the kinematics using the method described by Skals et al. [20]. Ground reaction forces were predicted by creating 25 contact points under each foot of the musculoskeletal model. Each contact point consisted of ve unilateral force actuators, which could generate a positive vertical force orthogonal to the ground, and static friction forces in the two horizontal directions using a friction coef cient of 0.8. In addition, to compensate for the sole thick- ness of the running shoes, a 25 mm height activation offset threshold was added to the musculoskeletal model. Each runner provided 24 running strides, giving 576 running strides in total (=24 runners 4 strides 2 legs 3 speeds ) with an estimate of patellar and Achilles tendon

friction coef cient of 0.8. In addition, to compensate for the sole thick- ness of the running shoes, a 25 mm height activation offset threshold was added to the musculoskeletal model. Each runner provided 24 running strides, giving 576 running strides in total (=24 runners 4 strides 2 legs 3 speeds ) with an estimate of patellar and Achilles tendon force. Each of the four strides was paired with the observation from Garmin closest in time to this measure.

Sensors2021,21, 7418 4 of 12 2.5. Statistics Data were analyzed using a mixed-effects multiple regression model, which was used to model the association between the estimated forces in anatomical structures and the training load variables during the xed running speeds. In the mixed model, runner- speci c random effects are used to account for possible correlations between repeated observations for each runner at different running speeds. The response variable is either the estimated peak force or impulse per stride in either the patellar tendon or Achilles tendon. The predictor variables are the different training load variables from Garmin (speed, ground contact time, vertical oscillation, and cadence) and anthropometric variables of the runners, including body mass (kilogram), sex, knee height (centimeter), ankle height (centimeter), shoe sole height (centimeter) and body height (centimeter) or the traditionally used running distance (approximated by distance per stride). Once a model is established for the relation between a response variable and the predictor variables, an algorithm for predicting the response variable is directly obtained from the model prediction equation. For each of the four response variables, three models were obtained. The rst model is based on the “distance” as input, which is the common way in the literature of estimating cumulative load and serves as a control [10]. The second model is the “best tted”, which considers all available predictors and serves as a benchmark for the predictive performance and should be used when possible (See Tables). The third model is the “practically feasible” which considers only predictors, which may be assessed validly by runners themselves. The reason for tting a best- tted and practically feasible model stems from our expe- rience from previous studies [7,17,22]. Some anthropometric measurements are dif cult to collect in a large-scale project (ankle and knee height), in contrast to other measures such as running measures, sex, and body height, which may be assessed validly by runners themselves. Hence, for the best- tted model, all 10 variables were available, while the practically feasible disregarded knee, ankle, and shoe sole height and the distance model only used the distance per stride. The

large-scale project (ankle and knee height), in contrast to other measures such as running measures, sex, and body height, which may be assessed validly by runners themselves. Hence, for the best- tted model, all 10 variables were available, while the practically feasible disregarded knee, ankle, and shoe sole height and the distance model only used the distance per stride. The best tted, as well as the practically feasible models, were identi ed by tting all potential combinations of predictors (best tted model:2 10 = 1024 and practically feasible model:2 6 = 64). For each response variable, the combination of predictors that minimized prediction error was selected. The prediction error (PE; Equation (1)) was computed for each combination of predictors using a 5-fold cross-validation approach [23,24] comparing the difference between the tendon force estimated using AnyBody and the tendon force predicted by the algorithm given by the considered combination of predictors. PE= q [observed predicted] 2 (1) A relative proportion of prediction error (PPE; Equation (2)) was developed to get an impression of the size of error within and between the structure-speci c forces: PPE= prediction error(N) mean structurespeci f ic(N) 100 (2) The statistical analyses were performed in R. ThePEandPPEwere also used to compare the best tted and practically feasible models with the distance-based model. In addition, we estimated the Pseudo R-Squared value based on the approach described by Nakagawa and Schielzeth [25].

Sensors2021,21, 7418 5 of 12 Table 1.The predictive algorithms of the patellar and Achilles tendon peak force during running based on outdoor measurable features. Structure and Model Garmin Measurable Variables Measurable Variables by Runners Accuracy Intercept (N) Stride Length (cm) Speed (km/h) Ground Contact Time Length (ms) Vertical Oscilation (mm) Cadence (Step/min) Body Mass (kg) Sex (1 = Male) Knee Height (cm) Ankle Height (cm) Shoe Sole Height (cm) Body Height (cm) Standard Deviation between Runners (N) Standard Deviation within Runners (N) Prediction Error (N) Proportion of Prediction Error (%) Pseudo R-Squared (Fixed Effects) Achilles tendon peak load Distance algorithm 487 [480] 4 [0.33] 831 1317 1532 30 0.1 Practically feasible algorithm 9928 [6853] 339 ** [29] 292 [662] 60 [39] 1267 837 1406 27 0.17 Best tted algorithm 7189 [5530] 288 ** [33] 31 ** [8] 32 * [24] 33 [104] 737 ** [214] 1152 ** [201] 78 [57] 690 833 993 19 0.54 Patellar tendon peak load Distance algorithm 4102 [277] 1.01 [0.19] 475 778 934 18 0.02 Practically feasible algorithm 1615 [4731] 141 ** [25] 46.56 * [18] 16 [22] 484 [373] 23 [31] 641 469 743 14 0.32 Best tted algorithm 1228 [5007] 147 ** [25] 53 ** [19] 10 [23] 650¤ [363] 140 [93] 256 [173] 175 [172] 90 ¤ [51] 597 469 678 13 0.39 ¤ indicates thep-value for the variable to be between 0.05 and 0.1; * indicates thep-value for the variable to be less than 0.05; ** indicates thep-value for the variable to be less than 0.01; [] indicates the standard error of the estimate. Table 2.The predictive algorithms of the patellar and Achilles tendon impulse load during running based on outdoor measurable features. Structure and Model Garmin Measurable Variables Measurable Variables by Runners Accuracy Intercept (kN) Stride Length (cm) Speed (km/h) Ground Contact Time Length (ms) Vertical Oscilation (mm) Cadence (Step/min) Body Mass (kg) Sex (1 = Male) Knee Height (cm) Ankle Height (cm) Shoe Sole Height (cm) Body Height (cm) Standard Deviation between Runners (N) Standard Deviation within Runners (N) Prediction Error (N) Proportion of Prediction Error (%) Pseudo R-Squared

This study investigates algorithms for predicting tendon forces during running using easily accessible data.