Changes in running economy and running technique following 6 months of running with and without wearable‐based real‐time feedback

An increasing number of commercially available wearables provide real‐time feedback on running biomechanics with the aim to reduce injury risk or improve performance.


| INTRODUCTION
Wearable technology is highly popular among runners, with ~90% of the runners in the United States using a Global Positioning System (GPS)-equipped watch or some other wearable device in 2017. 1 Similarly, research among Dutch marathon runners showed that 86% used at least one wearable during their preparation for the marathon in 2014. 2 Runners often use wearables to quantify various running variables in the hope that this can help them reduce injury risk and improve their performance. 3,4ariables typically measured by wearables include distance, speed, elapsed time, and increasingly often also (relatively simple) biomechanical variables such as cadence and contact time due to their reported relation with injury risk and running economy in some studies. 3,5,6For example, a prospective cohort study found a lower duty factor to be associated with a higher overall risk of running injuries after adjusting for running speed. 6While duty factor was not associated with running economy in a recent systematic review with meta-analysis, a higher step frequency was found to be correlated with a better running economy. 7espite the popularity of wearables, it remains largely unknown if they are effective at modifying biomechanical risk factors for running injuries when used outside of a controlled study setting.Specifically, while several studies with commercially available wearables (e.g., Garmin watch) have shown that real-time wearable-based feedback can be used to modify running biomechanics associated with injury risk, [8][9][10] these studies often also provided instructions on how to implement the wearable-based feedback in line with motor learning principles.Willy and colleagues 8 for instance instructed individuals to view the real-time feedback only on specific runs to aid internalization of the newly trained technique.However, such instructions are typically not provided when runners use commercially available wearables outside of a study setting, which may therefore reduce the effectiveness of these wearables.Similarly, feedback can be provided in different ways (e.g., auditive, visual) and at different frequencies, all of which may influence the effectiveness of the feedback to modify running technique. 3t also remains unknown if changes in running technique pursued by wearables are effective at improving running economy and by extension running performance, with only one study investigating the effect of wearable technology on running economy.Specifically, Hafer et al. 11 showed that real-time feedback on cadence via a metronome or music-app did not alter running economy, despite alterations in cadence.This finding agrees with lab-based studies were individuals received real-time feedback to modify their running technique, [12][13][14] and studies where individuals were provided with instructions and/or feedback to modify their technique by a trainer, [15][16][17][18] with both study designs typically reporting no improvements in running economy following the intervention.However, two lab-based studies did report improvements in running economy following cadence retraining among individuals whose cadence was (likely) lower than their most economical cadence. 19,20This suggests that gait retraining may only improve running economy for individuals whose self-selected technique deviates substantially from their most economical technique (which includes most novice runners according to 21 ), while it may reduce running economy for individuals who have already self-optimized their technique.Indeed, the only study reporting a decreased running economy following gait retraining was conducted among sub-elite triathletes, 22 who may already have optimized their technique.Another reason for the lack of improvements in running economy following gait retraining in most previous studies is that their duration was rather short.6][17][18] The duration of these studies may have been too short to see beneficial effects of changes in running technique on running economy.In support of this, 14 weeks of training to change from a rearfoot strike to a more midfoot/forefoot strike worsened running economy after 2 weeks, but resulted in no change in running economy on the long-term (14 weeks) (16).The authors provided evidence for the need of a gradual long-term intervention to overcome stability losses and associated higher energy costs associated with altered running technique.Longer time periods (e.g., >3.5 months) may therefore be required for alterations in running technique to translate into an improved running economy.Further support for the need of longer training periods is provided by a study were tibial-mounted accelerometer feedback did not significantly change running economy after 3 weeks of training, but a non-significant trend towards improved running economy was present at the 1-month follow-up. 12Finally, some findings indicate that running economy may mostly improve at the typical training speed, 23 yet most studies assessed running economy at the same fixed-speed for all individuals, which may have limited their ability to detect effects.
Overall, these findings highlight the need to investigate whether (A) commercially available wearables can effectively modify running biomechanics associated with lower injury risk, and (B) changes in running biomechanics pursued by wearables can improve (or maintain) running economy when measured over a longer (>3.5 months) time period, and at the participant's typical training speed.The primary aim of this study is therefore to investigate if real-time feedback on spatiotemporal metrics by a commercially available wearable is effective at modifying running biomechanics and running economy.We hypothesize that the wearable will modify running biomechanics towards a higher step frequency, which in turn is expected to be accompanied by alterations in kinematics such as a smaller peak knee flexion and decreased hip adduction during stance. 24The hypothesis that the wearable will modify running biomechanics towards a higher step frequency is based on two findings.First, knee (and specifically patellofemoral) injuries are the most common injuries among novice runners. 25Because several studies have found patellofemoral loading to decrease with increases in step frequencies, 26,27 we anticipate that the wearable will also aim to increase step frequency to reduce patellofemoral loading.Second, studies that have investigated running economy across different step frequencies indicate that novice runners often select a step frequency that is slightly lower than their theoretical most economical step frequency, 21 and we therefore expect the wearable to increase step frequency for most runners.Additionally, we hypothesize that changes in running biomechanics due to the wearable feedback will lead to larger improvements in running economy compared to the control group.As a secondary aim, we also explored associations between changes in running biomechanics and changes in running economy to gain further insight into possible biomechanical alterations that may contribute to changes in running economy.

| General study design
This study is part of a larger randomized controlled trial among 220 individuals that investigated the effects of real-time biomechanical and relative speed feedback on running injuries and running performance. 28A random subgroup of 40 participants (20 from each group) reported to the lab pre and post a ~6 month intervention to investigate changes in running biomechanics and running economy in controlled conditions.A stratified randomization by sex was done by generating random numbers per group using Rando mizer.org and approaching the participants with the assigned number until the required sample size was met.
All participants were instructed to avoid strenuous activity for 36 h, alcohol for 24 h, caffeine for 6 h, and a heavy meal 1 h before the session.When entering the lab, anthropometric measurements were taken using standardized procedures.The participants were then equipped with retroreflective markers for running experiments.After subject calibration and a familiarization period, the participants completed short (4 min) runs while ground reaction forces, full-body kinematics and respiratory gasses were collected.

| Participants
A total of 40 participants (21 male, 19 female, mean ± SD age 37.8 ± 11.5 years, body height 173.7 ± 7.9 cm; body mass 72.4 ± 10.4 kg; weekly running distance 25.5 ± 10.6 km; runs per week 3.0 ± 0.5 times; running experience 8.0 ± 7.4 years) that were a randomly selected subsample of a study investigating the effects of real-time feedback on running injuries and running performance volunteered to participate in the study.All participants were free of any moderate (for previous 3 months) or minor (for previous 1 month) musculoskeletal injuries, were comfortable with treadmill running (mean self-reported comfort from 0 to 10 was 8.7 ± 1.4), had a body mass index (BMI) of <27.5, and were aged 18-65 at the first test.Moderate injuries were defined as injuries that required at least 1 week of rest or reduced training volume, whereas minor injuries were defined as small complaints for ≥2 days in a row.The participants were also required to be self-identified novice/recreational runners, that were running maximally 40 km per week at the start of the study.The study was approved by the local ethics committee (nr.NL72989.068.20), was conducted according to the Declaration of Helsinki, and all participants signed informed consent prior to the measurements.The same participants underwent the pre, and post-intervention investigation and all running experiments were performed at approximately the same time of the day in similar conditions (mean ± SD temperature 20.7 ± 0.4 vs. 20.3 ± 1.1°), with each participant being instructed to wear their own and similar running shoes in the pre and post test.

| Intervention
All participants (i.e., both groups) were provided with (prototype) instrumented insoles (ARION, ATO-Gear, Eindhoven, the Netherlands) that could connect with a (beta) application each participant installed on their mobile phone.The participants were randomly assigned to an intervention group that received real-time feedback on spatiotemporal parameters and the relative speed of their run, or a control that did not receive real-time feedback (i.e., they were asked to turn off the real-time feedback in the wearable application).Group assignment was determined using an online research randomizer (Resea rchRa ndomi zer.org) by one of the researchers, with stratification based on sex.Each participant subsequently specified the assigned number in the beta-app created for the study to ensure anonymity of the collected data for the remaining study.Due to logistical reasons it was not possible to blind the researcher that performed the data analysis to the group allocation.
Both the intervention and control group were provided with real-time feedback not related to spatiotemporal parameters or relative speed, such as the running distance, absolute running speed, and duration.Participants in both groups could also review a summary of the recorded biomechanical metrics (e.g., average cadence) after each training session in the app.We chose to also provide participants in the control group with a summary of the recorded metrics as these metrics are also often available when running with other wearable technology. 3,29articipants in both groups were provided with general training guidelines as detailed in Data S1.
Overall, the wearable aimed to modify running technique and training intensity by real-time feedback with the aim to reduce injury risk and improve running performance.To this purpose, the application would first request users to run a baseline run (i.e., a run with no feedback) at their comfortable running speed as determined using the talk-test.The wearable then determined the comfortable running speed, and spatiotemporal metrics such as cadence and footstrike (FS) index from the pressure data as described and validated previously. 30These metrics were used as inputs to an algorithm that determined the goal for the next session (i.e., cadence, FS or running speed), along with the target zones as detailed in. 28Briefly, the wearable used correlations reported in the literature (e.g., 24 ) to infer relative loading of the foot/ankle and knee/upper leg.For example, a relatively low step frequency combined with a very pronounced heel strike was assumed to result in a relatively higher load at the knee than at the Achilles tendon or foot. 24Conversely, a relatively high step frequency combined with a very pronounced forefoot strike was assumed to result in a relatively higher load at the Achilles tendon or foot than at the knee.This relative load inferred from spatiotemporal metrics, was used with information about prior injuries, as input to an algorithm to generate individual targets zones for real-time feedback.When there was no current injury or pain, the wearable used the relative loading inferred from spatiotemporal metrics and literature to determine a target zone with the goal to gradually reduce the loading of the body part with the highest load.Fixed thresholds (i.e., independent of speed or slope) for cadence and FS index were used within the algorithm to this purpose.For example, a cadence of <170 steps per minute combined with a FS index of <33% may be used as cut-off for the algorithm to decide the runner could benefit from increases in cadence or FS modification to reduce the load at the knee.Note that the cut-off values mentioned here are solely examples as the exact thresholds are not publicly available.When a current injury/discomfort was specified in the app, the wearable attempted to specifically reduce loading on this body segment.For example, the real-time feedback attempted to reduce relative load on the knee by gradually increasing step frequency 24 relative to the individuals baseline.The focus for a particular session was prioritized according to a hierarchy with cadence selected ahead of FS, and intensity sessions recommended only after successfully completing sufficient technique sessions.Depending on the direction of the change required as deemed by the algorithm, the target zone was set by increasing or decreasing the baseline value by 5% and setting an upper and lower range around this value.A 5% value was used to ensure changes were made gradually over time.
As a secondary goal, the wearable also aimed to improve running economy by means of cadence and FS modifications.For example, the increase in step frequency for a runner with a low step frequency during the baseline run may potentially improve running economy by bringing this runner closer to his/her theoretical most economical step frequency.Conversely, if a runner adopted a very high step frequency during the baseline run, the wearable attempted to reduce step frequency based on the assumption that this may improve running economy by reducing internal mechanical work to displace the legs relative to the body center of mass.For all technique sessions, users were given instructions to run at a speed at which they could comfortably talk.If a user ran >5% faster than their baseline run during a technique session, the wearable would inform the user about this to maintain speed within a safe range.
The wearable also included periodic sessions with a focus on relative running speed that could either be a "stable speed run" that focused on maintaining an intensity within ±5% of the runners' comfortable baseline running speed (80% of the sessions), or intensity session, which involved higher speed running within ±20% of the baseline running speed (20% of the sessions).
Real-time feedback was provided via a mobile phone using auditive instructions when the moving average over the last 20 steps was 5% lower or higher than the target value.Auditive feedback was provided again after another 20 steps if the moving average was still 5% higher or lower than the target value and this process was repeated for the complete duration of the session.The feedback was deemed to be successfully followed if the participant remained within the target-zone for the target parameter for more than 80% of the running steps performed during the session, thereby allowing the participant to progress to sessions with feedback on other metrics.Note that the application did not provide a detailed training program and runners could therefore self determine when to run and how long they wanted to run.
Breen and colleagues 31 showed that just three sessions of coaching over a 6 week period resulted in significant changes to running technique when measured without feedback at the 6-week measurement.This suggests that at least three sessions of feedback could be sufficient to cause long-term changes to running technique.To be included in the intervention group, we therefore a priori decided that participants were required to have ran at least three sessions with real-time feedback, and to additionally have received real-time feedback during at least 60% of their sessions to ensure sufficient exposure to feedback relative to no feedback when a larger number of sessions was performed.To minimize exposure in the control group to real-time feedback, these participants were allowed to have run maximally three sessions with real-time feedback.Individuals that ran more than three sessions with feedback, but did not met the 60% threshold were not included in the analyses.However, this was not the case for any participant.To check the sensitivity of our decision on the results, we also performed an analysis where individuals in the intervention group were required to have run at least 10 sessions with real-time feedback, and to additionally have received real-time feedback during at least 60% of their sessions.This resulted in the exclusion of one participant from the analysis.

| Assessment of running technique and running economy
The computer assisted rehabilitation environment (CAREN, Motek, the Netherlands) system combines an instrumented split-belt treadmill (belt length and width 2.15 × 0.5 m, 6.28 kW motor per belt, 60 Hz belt speed update frequency and 0-18 km h −1 speed range) with a 12-camera three-dimensional motion capture system (VICON NEXUS v2.7, Oxford Metrics Group, Oxford, UK, 100 Hz) and was used to determine kinetic and kinematic outcomes during running at various speeds (Figure 1).
After calibration of the systems, 46 retroreflective skin markers with a diameter of 14 mm were attached to the skin with double-sided tape using a modified full-body marker (Human Body Model v2) set. 32In line with the International Society of Biomechanics (ISB) suggestions, each participant was asked to perform basic motion tasks of the hip and knee joint to determine the "functional" axis of rotation.After the subject calibration, the participants performed an 8-min familiarization trial. 33The speed of this familiarization was set to the participant's self-reported typical training speed.After 4 min on their typical training speed, the participants continued in self-paced modus, whereby the treadmill automatically adjusted the belt speed based on the position of the participant (as determined by the pelvic markers) in relation to the belt. 34The sensitivity and acceleration of the selfpaced algorithm were set to 1.5 and 1.5 ms 2 , respectively as pilot experiments indicated this better matched realworld accelerations compared to the default settings.The average running speed over the last minute of self-paced running was taken as comfortable running speed.After the familiarization, participants ran again at their comfortable running speed, 10% faster, and at a fixed speed of 2.78 and 3.33 ms −1 , with the order of the conditions being randomized.The duration of all conditions was 4 min, and the last minute was used for gas exchange and biomechanical data analysis.During the post-test, the participants' comfortable running speed was re-assessed using the same procedure and same starting speed.The participant then ran at the same speeds as during the pretest.Running economy and biomechanics were assessed at all four speeds.We assessed these outcomes at the participants' comfortable speed because running economy may mostly improve at the typical training speed. 23I G U R E 1 CAREN system with simulated virtual forest environment prior to initiation of running.The participant is also wearing the ARION wearable (note the green lights on the inertial measurement unit attached to the side of the shoe).
Moreover, the 10% faster speed was used as the wearable also incorporated interval sessions, which may contribute to improvements in running economy at higher speeds.The fixed speeds were additionally used to assess if the magnitude of changes observed were smaller than any changes observed at an individualized speed, and to allow between-subject comparisons without a confounding effect of running speed.
For respiratory gas analyses, participants wore a face mask (Hans Rudolp Inc, Shawnee, KS, USA) over the nose and mouth without detectable leakage.The mask was connected to a T-piece that was placed in a free airstream (200 L min −1 ).Respiratory gases were captured using a total-capture indirect calorimeter (Omnical, Maastricht Instruments, Maastricht, the Netherlands) 35 to investigate running economy at each speed.The system was calibrated automatically every 15-30 min using room air and a gas mixture of known composition.Steady-state VȮ 2 can typically be reached within approximately 2-3 min at each running speed and a duration of minimal 4 min was therefore used for each speed to ensure a 1-min steadystate period could be used for biomechanical and gas exchange data analysis.A steady-state VȮ 2 and VĊO 2 was visually confirmed and the trial duration was extended if no steady-state was achieved.Trials that yielded a respiratory exchange ratio (RER) >1.0 were excluded.The participants were allowed to take rest periods between trials when required and were instructed to run as if they were running outside and to focus on the simulated virtual forest environment (Figure 1).

| Biomechanical outcomes
Three-dimensional running kinematics were determined in real-time using a lower body and trunk musculoskeletal model (Human Body Model v2, consisting of nine rigid body segments, 21 degrees of freedom and 86 muscles) implemented in the D-flow software.The short computation time with real-time analysis results in negligible errors compared to unlimited computation time for the inverse kinematics analysis. 32Ground reaction forces and kinematics were filtered using a single-pass Butterworth filter with a low-pass cut-off of 20 Hz.A 20 Hz cut-off was chosen because most treadmill noise is above 20 Hz. 36 The filtered data were further analyzed using custom-written algorithms in MATLAB to extract and compute variables of interest.FS was identified when the vertical ground reaction force from force plate data exceeded 20 N and toe-off was identified when vertical ground reaction force dropped below 20 N. Spatiotemporal outcomes considered for analyses were ground contact time, flight time, duty factor, and step frequency.Ground contact time represented the time between foot contact and toe-off and flight time was the difference between stride time (i.e., the time from one foot contact to the next contact of the same foot) and ground contact time.Duty factor was determined as the ratio between flight time and stride time.Step frequency was taken as the number of steps per minute.FS angle was additionally determined as the angle between the ground and vector from the heel to toe (MT2) marker.
Positive values indicate a rearfoot strike while negative values indicate a forefoot strike.Sagittal-and frontal-plane pelvis, hip, knee and ankle angles were derived from the Human Body Model and custom-written MATLAB algorithms were used to time-normalize all kinematic data from 1% to 100% of the gait cycle.Results are only reported for the right leg as preliminary analyses revealed similar results for the left leg.

| Running economy
The rate of oxygen consumption (VȮ 2 ) and carbon dioxide (VĊO 2 ) production were measured continuously and computed at 5 s intervals throughout the running trials.VȮ 2 and VĊO 2 were subsequently used to determine substrate utilization using non-protein equations, 37 with energy cost being determined as the sum of fat and carbohydrate use.The energy cost was then expressed as J kg m −1 .

| Statistical analysis
The estimated marginal means from a repeated measures linear-mixed model fitted with a restricted maximum likelihood method (implemented in SPSS Version 25 IBM Corporation, Chicago, IL) were used to compare the change in running economy and spatiotemporal outcomes from pre to post measurements within each group and between groups.To this purpose, speed and group were specified as a fixed effect, the interaction between speed and group was included, and a random intercept and slope were modeled per subject.To maintain a familywise error of α = 0.05, the alpha level for rejecting the null-hypothesis of no change within each group or no difference between groups at each of the four speed (i.e., at comfortable, 10% faster, 2.78 and 3.33 ms −1 ) was corrected using the Bonferroni procedure, resulting in an effective alpha level of 0.05/4 = 0.0125.Normality of the residuals was assessed visually using Q-Q plots and histograms.Percentage changes were reported for all spatiotemporal and running economy outcomes by dividing the mean change by the pre-test mean, multiplied by 100.
The change in joint angles between the pre-and posttest, and any differences between groups in this change were assessed using a one-dimensional statistical parametric mapping (SPM) two-way repeated measures ANOVA implemented in Matlab.To maintain a family-wise error of α = 0.05, the Bonferroni procedure (p = 0.0125) was used to correct for the four speeds assessed per joint angle.We did not further correct for the multiple joints assessed as this would have yielded an overly conservative threshold, likely reducing the possibility to detect meaningful differences and thus increasing the risk of type II errors.
The overall correlation across all speeds between changes in spatiotemporal (repeated-measures) metrics and changes in running economy was determined using the "rmcorr" package in RStudio v4.2.3, 38 with 95% confidence intervals being computed using the bootstrap option.The assumption of fixed regression slopes across each speed within this package was checked visually by plotting linear regression lines with variable intercept and slopes.sessions with feedback because they did initially not deactivate the feedback feature that was on by default.Because the overall percentage of sessions ran with feedback was >60% for these individuals, we allocated them to the intervention group.This resulted in a final sample size of 9 for the intervention group and 13 for the control group in the analyses.The data presented for both the pre-and posttests in the remaining of this paper will reflect only the participants that participated in both test sessions.

| RESULTS
The average number of sessions performed, distance ran, and speed by participants in the control and intervention groups is reported in Table 1.

| Changes in running economy and comfortable running speed
In the pre-test, the RER was >1.0 for one participant during the comfortable running speed, for five participants during the 10% faster condition, and eight participants while running at 3.33 ms −1 , and these trials were therefore excluded from analysis.During the post-test, the RER was also >1.0 for one participant during the comfortable running speed, for six participants during the 10% faster condition, two participants during the 2.78 ms −1 condition, and three participants during 3.33 ms −1 .
There was a significant overall reduction in the energetic cost of running from the pre-to post-test in both groups when all speeds were combined (−0.26 ± 0.11 J•kg•m −1 , −6.14%, p = 0.02; −0.33 ± 0.12 J•kg•m −1 , −7.67%, p = 0.01 for the control and intervention groups, respectively).Similarly, after Bonferroni correction, the reduction remained significant for half of the running speeds in both groups (Table 2; Figure 3).The change in running economy did however not significantly differ between the groups overall (−0.07 ± 0.14 J•kg•m −1 , −1.54%, p = 0.63), nor at any of the individual speeds (Table 2; Figure 3).A sensitivity analysis whereby one participant in the intervention group that ran <10 sessions was excluded yielded similar results.Moreover, there were small, but nonsignificant significant correlations between the number of sessions or distance ran and the change in running economy overall (r = 0.17 and 0.11, respectively).
The participants' comfortable running speed did not significantly change from pre to post within each group, and there was also no significant difference in change between the groups (Table 2).

| Changes in spatiotemporal metrics and running biomechanics
There was a significant overall increase in contact time from the pre to post-test for the control, but not intervention group (3.65 ± 1.49 ms, 1.43%, p = 0.01; 2.72 ± 1.79 ms, 1.08%, p = 0.13 for the control and intervention groups, respectively).After Bonferroni correction, significant increases were observed for one running speed in the control group (Table 2).The change in contact time did not significantly differ between the groups overall (−0.93 ± 2.32 ms, −0.35%, p = 0.69), nor at any of the individual speeds (Table 2).There was a significant overall decrease in flight time from the pre-to post-test for the control, but not intervention group (−6.04 ± 2.56 ms, −6.25%, p = 0.02; −5.31 ± 3.07 ms, −5.59%, p = 0.08 for the control and intervention groups, respectively).After Bonferroni correction, no significant decreases were observed for individual running speeds (Table 2).The change in flight time did not significantly differ between the groups overall (0.73 ± 4.0 ms, 0.66%, p = 0.86), nor at any of the individual speeds (Table 2).There was a significant overall increase in duty factor from the pre-to post-test for both the control and intervention groups (0.81 ± 0.31%, 2.25%, p = 0.01; 0.74 ± 0.37%, 2.05%, p = 0.047 for the control and intervention groups, respectively).After Bonferroni correction, significant increases were observed for two of the individual running speeds (Table 2).The change in duty factor did not significantly differ between the groups overall (−0.08 ± 0.48%, −0.20%, p = 0.88), nor at any of the individual speeds (Table 2).There was no significant overall change in step frequency from the pre-to post-test for any of the groups (1.47 ± 1.18 steps•min −1 , 0.86%, p = 0.21; 1.74 ± 1.41 steps•min −1 , 1.00%, p = 0.22 for the control and intervention groups, respectively).Similarly, after Bonferroni correction, no significant changes were observed for any of the individual running speeds (Table 2).The change in step frequency did also not significantly differ between the groups overall (0.27 ± 1.84 steps•min −1 , 0.14%, p = 0.88), nor at any of the individual speeds (Table 2).FS angle did not significantly change for both groups overall (1.19 ± 2.02°, 26.5%, p = 0.17; 0.31 ± 2.42°, 4.14%, p = 0.26 for the control and intervention groups, respectively), and the change did also not differ between groups overall (−2.40 ± 3.15°, 22.4%, p = 0.46), nor at any of the individual speeds (Table 2).A sensitivity analysis whereby one participant in the intervention group that ran <10 sessions was excluded yielded similar results for all outcomes.The remaining results and discussion will therefore present the findings for the complete sample.Statistical parameter mapping showed no significant main effect for group and no interaction, thus demonstrating no differences in 3D running kinematics between the pre-and post-measurement within the control or intervention groups, nor any difference in the change in kinematics (Figure 4).

| Relation between changes in running economy and changes in spatiotemporal metrics or running biomechanics
Across all speeds and both groups combined, changes in flight time, duty factor or step frequency were not significantly associated with changes in running economy (Table 3).However, increases in contact time were significantly associated with reductions in the energetic cost of running (Table 3; Figure 2).

| DISCUSSION
The primary aim of this study was to investigate if real-time feedback on spatiotemporal metrics and relative speed by commercially available instrumented insoles is more effective at modifying running biomechanics and running economy than unsupervised running training without real-time feedback.Overall, our findings show that there were no significant differences in changes in running biomechanics or running economy between the group receiving real-time feedback and unsupervised running training when no feedback was provided to either group in controlled lab conditions.Specifically, both groups significantly reduced their energetic cost of running during the ~6 month intervention, but this reduction did not significantly differ between the groups.Similarly, while several spatiotemporal metrics significantly changed during the intervention period for both groups, the magnitude of the changes were minor and did not significantly differ between the groups.Finally, sagittal-or frontal-plane running kinematics did not significantly change within each group and the changes did also not differ between the groups.

| Running economy and comfortable running speed
The average improvements in running economy across all speeds were 6.1% and 7.7% for the control, and intervention  groups, respectively.4][45] and smallest worthwhile change of ~2.5% 43 reported for running economy in other studies and may therefore be considered clinically relevant.For example, such improvements would translate into a 3-4 min reduction in 10 km time for a recreational runner, running at 3 ms −1 (~56 min 10 km) at baseline. 46For both groups, the magnitude of the improvement was largest at the participants' comfortable running speed (Table 2).This finding is in line with previous studies that also show running economy improves mostly at the typical training speed, 23 and suggests future studies should consider assessing running economy at the participant's' typical/ comfortable running speed to maximize the chances of observing an effect of the intervention.In contrast to our hypothesis, running economy did however not improve significantly more in the group that received real-time wearable based feedback than the control group, although the magnitude of the change was slightly larger (~1.5%) in F I G U R E 3 Pre-and post-test values for running economy (A) and contact time (B) at the comfortable running speed for each group.Thick lines depict the mean effect, while thin lines depict a single individual.(C) depicts the relation between changes in contact time and running economy at each speed, with symbols representing the individuals grouped per speed, and the lines depicting the best fitted linear regression line.The variance in changes in running economy explained by changes in contact time (R 2 values) were typically low to moderate (14% for 2.78 ms −1 , 34% for 3.33 ms −1 , 33% for comfortable speed, and 34% for 10% faster).

F I G U R E 4
Pre and post sagittal-plane knee joint angle during the comfortable running speed in the control group (A), and intervention group (B).(C) depicts the difference in knee joint angle from pre to post measurement for each group.Statistical parameter mapping showed no differences within each group or between groups.

Correlation coefficient (95% CI) a
Step frequency 0.15 (−0.20 to 0. intervention group (Table 2, Figure 3).The difference of 1.5% is however within the typical error for measuring running economy reported in most studies (~2.5%); [43][44][45] and the improvement in both groups can therefore be considered equivalent.However, the intervention group achieved this with a smaller total and weekly distance (Table 1).Notably, for both groups, the lab measurements were performed without real-time feedback and running economy may have been altered during the in-field sessions with real-time feedback.
One could expect that increases in running economy would also lead to increases in the comfortable running speed, as this would allow individuals to run faster for a given (perceived) energetic cost.However, there was no significant change in the comfortable running speed in either group, nor in both groups combined over the study period (Table 1).The lack of change in comfortable running speed is in agreement with a previous study that also found no change in the preferred running speed (assessed overground) following 12 weeks of running training. 47The apparent discrepancy between changes in running economy and comfortable running speed suggests that comfortable running speed is not solely determined by the energetic cost (i.e., running economy), but also by other factors such as afferent feedback on mechanical load (e.g., via Golgi tendon organs) and optical flow.Related to this last point, our findings indicate that the comfortable running speed assessed in a lab may not always agree with the typical running speed in-field as the self-determined comfortable running speed in the lab was significantly higher than the average running speed in-field (by 0.24 ± 0.43 ms −1 ; p = 0.003; Figure S1).This is in contrast to the findings of a systematic review which showed runners typically run slower on the treadmill as compared to overground. 48The differences in comfortable running speed between the lab and field and compared to previous studies may reflect differences in optical flow between treadmill running in a virtual reality and overground running.Specifically, novice runners perceive the running speed on the virtual reality to be slower than the actual running speed 49 and may therefore have increased their speed to match the optical flow from the virtual environment with their overground running optical flow.This higher speed resulted in a RER >1 for three individuals during their self-determined comfortable speed in the pre-test (note that only one of these individuals also completed the post assessment), which is reflective of a high-intensity and non-sustainable/uncomfortable effort.In other words, although we instructed these individuals to run at a comfortable running speed, the selected speed did not reflect a comfortable speed as judged from the RER.Interestingly, for one of these individuals, the comfortable running speed assessed in the lab did not substantially differ from the average running speed in-field as assessed with the wearable (3.23 vs. 3.16 ms −1 in the lab vs. in-field, respectively).This suggests that some individuals may be training structurally too fast, which could contribute to a higher injury risk by increasing fatigue, and depleting carbohydrate stores, 50 which in turn may increase the risk of e.g., bone injuries. 51This therefore may suggest the need for real-time feedback on an appropriate training intensity to optimize performance and reduce injury risk for some individuals.

| Running biomechanics
There were significant changes for some spatiotemporal metrics (contact time, fight time, duty factor) from the pre to post measurement within each group.However, the magnitude of the change was minor.For example, the average increase in contact time of 3.7 and 2.7 ms for the control and intervention groups, respectively, are smaller than the minimum detectable change for contact time reported in some studies. 29,52The (minor) changes in spatiotemporal metrics during the study period did not significantly differ between the groups (Table 2; Figure 4).Similarly, 3D kinematics did not significantly change within or between groups.These findings are in line with previous studies that found no significant/minor alterations in spatiotemporal metrics 47,53,54 or kinematics 54 following various (supervised) training programs where training intensity/duration or frequency was increased over a duration of 6-12 weeks.Overall, these findings therefore indicate that unsupervised running training resulted in similar biomechanical alterations as running with real-time feedback once feedback is removed.
The absence of significant differences between groups in running biomechanics can be due to several reasons.First, it could be hypothesized that the participants in the intervention group might not have complied with the provided feedback.However, the participants ran within the target zone set by the wearable for the majority of the session duration (Table S1; Figure 5), thus demonstrating compliance with the feedback and making this an unlikely primary reason for the absence of changes.A closer inspection of the target zones does however show these to be relatively similar to the self-selected gait for most of the participant's runs (Figure 5).This therefore suggests that the primary reason for an absence of differences is that the target zones set by the wearable may not have been different enough from the participant's self-selected gait to induce learned changes that persist after feedback is removed.Moreover, for sessions where the target zones were different to the participant's self-selected gait and where the participant also conformed to the feedback, the participant may have relied on the feedback to alter their technique and thus did not to alter their gait without feedback (e.g., in the lab).In partial support of this, some participants did seem to alter their cadence or FS index during some sessions with feedback, but their cadence or FS index returned to their "typical" values during runs without feedback (Figure 5A).Some participants may therefore have been dependent on feedback and may not have learned to alter their running biomechanics in the absence of feedback.A possible explanation for the absence of learning effects could be that there were insufficient sessions with a substantially different target zone to induce learning.Additionally, the structure by which different outcomes where provided with feedback could also have influenced the learning effect.Specifically, the app provided feedback on different outcomes (cadence, FS index, relative speed), and did so in a structure that mimics random practice from a motor learning perspective (Figure 5).Previous studies showed that a period of blocked practice prior to random practice increases the learning effect in more complex skills, 55 and the random F I G U R E 5 Cadence and footstrike index recorded with the wearable during the in-field sessions in two participants.When low and upper target zones are indicated (horizontal black lines), there was real-time feedback provided on that metric during the specific session.For example, in (B), session 10 was a session with real-time feedback on cadence, with an upper target zone of 162, and lower target one of 142 steps•min −1 .Similarly, session 12 was a session with real-time feedback on footstrike index, with an upper target zone of 33%, and lower target one of 0%.Shaded areas indicate a baseline run without any real-time feedback.Note that for the participant depicted in (A), the target zones for cadence and FS index are very similar to the typical variation seen in the first four baseline sessions (shaded areas), suggesting the absence of effects on running biomechanics could be due to the target zone not being different enough from the baseline variation to induce acute and thereby long-term changes in running technique.For (B), the cadence target zone for session 29 and 33 are higher than the typical cadence and the participant also adopted a slightly higher cadence during these sessions.Yet after these sessions, cadence returned to the typical value, thus suggesting no learning effect.Finally, in session 36 and 40, the FS index target zones were also higher than the typical values, but the participant did not run within the target zone, potentially also contributing to the absence of changes in running biomechanics with the wearable relative to the control group.feedback structure could therefore also have contributed to the absence of learning effects in some participants.Alternatively, the participants may have learned to use a different technique, but chose to not adopt the newly learned pattern.Recent evidence for example shows that runners can adopt a running technique to reduce impact forces 6 months post intervention, but do not adopt this technique, except when specifically asked to do so. 56Such an effect would question the usefulness of the feedback for reducing injury risk of improving performance when running without feedback.A final explanation for the absence of between-group differences is that the runners in the control group may also have self-optimized towards a more economical (e.g., longer contact time) or less injury prone running technique, which in turn masked any between group differences.Indeed, previous studies have also shown self-optimization of running biomechanics during running training programs. 40

| Correlation between changes in running economy and changes in spatiotemporal metrics
A secondary aim was to explore the relation between changes in the energetic cost of running and changes in running biomechanics.We decided to only investigate the association between spatiotemporal metrics and running economy due to the absence of changes in running kinematics.Moreover, we performed this analysis with the data from both groups combined due to the small sample sizes in each group.Among the spatiotemporal metrics, only increases in contact time were significantly associated with improvements in running economy (Figure 3; Table 3).In line with our finding, a previous study also reported that 8 weeks of combined continuous, interval and strength training among recreational runners significantly improved running economy and prolonged ground contact time. 39Although most runners naturally select contact times that are close to their optimum value, one study found that runners often select slightly shorter contact times than optimal. 57Contact time may therefore have increased due to self-optimization, whereby a longer contact time may reduce lower-limb fascicle shortening velocities and hereby improve running economy (e.g., 58 ).Nevertheless, the strength of association between changes in running economy and changes in contact time was typically small to moderate within each speed (R 2 = 0.11-0.34; Figure 3), suggesting that other mechanisms (e.g., tendon stiffness) also contributed to changes in running economy.
In this context, it is also important to discuss if an increased familiarization with treadmill running may have contributed to the better running economy.Specifically, it may be hypothesized that the participants were more comfortable with treadmill running during the post test, which in turn may also have contributed to the improved running economy.However, we believe the influence of familiarization on the improved running economy to be small for the following reasons.First, the participants were already very comfortable with treadmill running prior to the study (mean self-reported comfort from 0 to 10 was 8.7 ± 1.4), and were given an additional 8 min of familiarization prior to data collection in line with recent guidelines. 33Moreover, unfamiliarity with treadmill running would mostly be reflected by a higher cadence, 33 yet cadence did not differ between the two measurements and was rather higher than lower in the post test (Table 2).

| Self-reported vs. measured running distance
An interesting observation was that the self-reported running distance at the time point of recruitment was considerably higher than the running distance recorded with the wearable during the study (25.8 ± 11.4 km•w −1 for the participants that completed both tests vs. 12-8 km•w −1 as recorded by the wearable for the control and intervention groups, respectively).There are several potential explanations for this observation.First, it could suggest that the participants ran only ~50% of their sessions with the app, or it could reflect inaccurate distance recording by the app.However, the larger project 28 found that the overall incidence of 9.05 injuries per 1000 h for injuries >7 days was only slightly higher than the incidence among recreational runners reported by a systematic review (7.7 injuries per 1000 h) in recreational runners, 59 thus suggesting that participants did run most of their sessions with the app because the injury rate per 1000 h would otherwise be expected to be considerably higher than published values.Weekly training distance obtained from self-reports has also been shown to agree closely with weekly running distance from GPSs, 60 although there was considerably variability in the agreement.This therefore suggests that this may also not be a primary factor explaining this difference.Rather, the baseline questionnaire was done approximately 2 months prior to the commencement of the study and participants may have changed their running routine during the course of the study.For example, the baseline questionnaire was completed in January 2021 for most participants.This was during or just after a COVID-19 lockdown during which participants may have had more time to run.
When interpreting the improvements in running omy observed for both groups (Table 2) in relation to the reduction in weekly training volume, they suggest that the improved running economy may reflect the higher accumulated distance during the ~6 month intervention period.

| Limitations
There are several limitations that should be kept in mind when interpreting the findings of this study.First, about half the sample size dropped out due to various reasons (Figure 2) and this reduced our statistical power.Nevertheless, the combination of multiple speeds for statistical analysis partly compensated for this lower power.Related to this, some participants changed group allocation because they did not use the correct settings in the app.We analyzed participants according to their final group (as-treated) and this therefore reduced the effect of the randomization.A second limitation is that we did not analyze transverse-plane kinematics, nor kinetics.However, given that there were no significant changes in the sagittal-or frontalplane kinematics and typically minimal changes in spatiotemporal outcomes, we considered it unlikely that there were relevant changes in these outcomes.A third limitation is that we did not standardize shoe wear in the lab experiments, and alterations in shoe wear could therefore have influenced our findings.Nevertheless, we instructed participants to wear the same pair of shoes in the pre and post measurement, and thus expect this effect to be small.A fourth consideration is that the control group was also provided with real-time feedback on running distance, absolute running speed and duration, and they could review a summary of the recorded biomechanical metrics (e.g., average cadence) after each training session in the app.While this may have reduced the contrast between the groups, we purposely chose to also provide participants in the control group with this feedback as this is also often available when running with other wearable technology and thus improves the ecological validity.Moreover, the larger study 28 showed that participants did not know how to use the summary of the biomechanical metrics to alter their gait, thus making it unlikely that this could have contributed to smaller differences.A fifth limitation is that the weekly running distance was ~30% higher in the control group than in the intervention group (Table 1).While it could therefore be argued that the app may have contributed to an equal reduction in running economy despite a smaller distance, further research is required to investigate this notion.A final limitation is that the cut-off values used for defining a high or low step frequency or FS were independent of running speed or slope within the version of the application used for this study.As cadence and FS index are speed dependent, slower runners may therefore have been more likely to receive feedback to increase their cadence and FS index, while faster runners may have been more likely to receive feedback suggesting to decrease these metrics.

| Perspective
Our findings show that both the group that received realtime feedback, and the group that received no real-time feedback significantly reduced their energetic cost of running during the ~6 month intervention.When measured without feedback in controlled lab conditions, there was no significant difference between the groups in the change in running economy.Similarly, while some spatiotemporal metrics significantly changed during the intervention period there was no significant difference in the change between groups.There were also no significant differences within or between groups in 3D running kinematics.In the present study, real-time feedback from the ARION wearable did therefore not result in significant differences in running biomechanics, or improve running economy more compared to unsupervised running training when assessed in controlled conditions while running without feedback.
When the findings of an absence of effect of the wearable on running biomechanics is interpreted in relation to the significant reduction found in injury rate with this wearable in a recent study, 28 it can be speculated that the reduced injury risk may result from acute changes in biomechanics for some participants while running with in-field real-time feedback, and potentially from more controlled running speeds with real-time feedback on relative intensity.This therefore also implies that injury risk may not decrease when continuing to run without the real-time feedback due to the absence of an apparent learning effect.Because the effectiveness of a wearable depends on various factors such as the feedback frequency, content, and modality, 3 and duration and number of sessions in which feedback is provided, further research is required to investigate the influence of these factors and their interaction on the effectiveness of wearable-based feedback.

Figure 2
Figure 2 depicts the flow chart of the included participants.Ten participants dropped out in the intervention group and eight in the control group, for reasons depicted in Figure 2. A total of five participants changed group allocation during the study.Specifically, three individuals in the intervention group ran <3 sessions with feedback and were therefore allocated to the control group.Two individuals that we allocated to the control group ran ≥3 Descriptive statistics from the pre and post-test represent mean ± SD, whereas the within group change and between group difference in change are presented as mean ± SE.Significant changes within groups, or significant differences in the change between group after Bonferroni correction (p < 0.0125) are highlighted in bold.a The difference in change score is presented as intervention relative to control group.T A B L E 2 (Continued) reflect the correlation after removal of three outliers with very large reductions in the energetic cost of running.Prior to outlier removal, the correlations were 0.18 (−0.03 to 0.41) for step frequency, −0.46 (−0.63 to −0.22) for contact time, 0.15 (−0.19 to 0.37) for flight time, and − 0.22 (−0.40 to 0.12) for duty factor, respectively.*Significant after Bonferroni correction (p < 0.0125).
Mean ± SD descriptive statistics for the intervention and control groups.Descriptive and inferential outcomes for the pre and post measurement for both groups.
T A B L E 1Abbreviation: SE, standard error.aMedian and interquartile range.T A B L E 2