The search for procedures allowing quantitative assessments of the cognitive performance of players in real situations represents a challenge for sport scientists and coaches (1). In the case of soccer, a game characterized by a huge variety of perceptual and decisional processes (2), there is no standardized test that measures the cognitive performance in field situations (3), which is in sharp contrast to the existence of several standardized field tests allowing the assessment of physical performance parameters (4).
In a recent study (5), we introduced a quantitative test to assess cognitive-motor performance (CMP). In the context of soccer, CMP is defined as the ability “to quickly gather game-relevant information and use this information to adequately execute a certain motor task.” Using a simple passing test, we observed that with increasing age of the players, CMP improved linearly with average gains of 4 cm (passing accuracy), 2.3 km·h−1 (passing speed [PS]), and 30 ms (response time [RT]) per year of age. Importantly, coaches who were asked to evaluate CMP performance of the participating players differed considerably in their evaluation, and only the mean value of all coaches came close to the objectively assessed CMP values. This shows that it is extremely difficult to judge CMP performance without any objective and reliable measures.
In the present study, we followed an approach based on two concepts, namely augmented feedback (for a review, see (6)) and perceptual learning (for a review, see (7–9)). Augmented-feedback learning is well known in sport sciences: it refers to the positive role of external feedback provided to participants when learning new skills or improving the efficiency of learned skills (6). Here, external feedback corresponds to any feedback provided in addition to the feedback naturally transferred by sensory receptors during task execution (10). In visual neuroscience, perceptual learning refers to the “long-term performance increase resulting from visual perceptual experience,” even for basic visual functions like visual acuity (9). This visual perceptual experience can be acquired through repetitive exposure to specific stimuli. A nice example is that of a radiologist who can identify cancer in an x-ray picture that untrained observers cannot see (9). Similarly in soccer, the repetitive exposure to specific visual stimuli allows goalkeepers to improve their prediction of penalty directions (11). In this study, goalkeepers followed home training using repetitive stimuli (temporally occluded videos of penalty kicks filmed from the perspective of the goalkeeper) and were asked to predict the direction of the ball: goalkeepers who received a feedback during training improved their anticipation skills (assessed through verbal responses of goalkeepers following presentation of occluded videos of penalty kicks) significantly more than the placebo (no feedback provided during training) and control groups.
The interaction of the underlying sensory–perceptual mechanisms with the motor execution level relies on a brain circuitry which is involved in both trial-and-error learning and augmented feedback learning (12).
In this study, we tested the hypothesis that augmented feedback training can improve both perceptual–cognitive and motor skills during a passing situation in soccer, namely, the capacity to accurately anticipate the visual consequences of an action and/or the quality of the movement itself. This led us to design a passing situation which is more demanding than the one we previously tested (which included only static visual targets and distractors, see (5)) and closer to a typical game situation. Indeed, passing the ball to a moving partner requires the cognitive ability to correctly anticipate the future position of the partner and the motor ability to adequately adjust the pass force and direction following this estimation. To this purpose, we monitored the improvements in passing performance of three groups of young elite soccer players, each receiving different types of training or feedback.
Twenty-seven elite young soccer players (males, U14 and U15 categories) participated in this study. They belonged to the same elite youth soccer academy and were playing at a Swiss-national level (highest level in Switzerland). They had participated in the sport within a club for a minimum of 5 yr and were training for at least 7 h·wk−1 (divided into four to five sessions per week) in addition to the weekly competition yielding a minimum of 8.3 h of soccer-specific practice per week. For the purpose of the study, they were divided into three groups of nine players: the augmented-feedback training group (AF group, age = 14.6 ± 0.4 yr, whole soccer practice experience = 6.4 ± 0.9 yr, including 4.1 ± 0.8 yr at the elite level), the no-feedback training group (NF group, age = 14.3 ± 0.6 yr, whole soccer practice experience = 6.7 ± 0.9 yr, including 3.9 ± 0.9 yr at the elite level), and the control group (CON group, age = 14.3 ± 0.8 yr, whole soccer practice experience = 6.6 ± 0.7 yr, including 3.9 ± 0.8 yr at the elite level). These players were trained by five expert coaches (including one goalkeepers’ coach). All participants (coaches, players, and players’ parents) provided informed consent, and the research procedures were approved by the local ethics committee. The experiments took place in a covered hall within the National Youth Sports Centre of Tenero (Switzerland).
Six coaches (one head coach of the academy and the five coaches mentioned above) participated in the study by providing their judgments about the performance level of the young players under their responsibility (see Procedure section). The coaches were experienced (10.5 ± 7.9 yr of coaching practice, including 5.8 ± 3.7 yr at the elite level) and certified trainers. They hold Union of European Football Associations-A (n = 3) and Union of European Football Associations-B (n = 3) licenses.
The COGNIFOOT system (patent pending at the Swiss Federal Institute of Intellectual Property under the reference CH00215/16) is a real-time high-technology system combining a visual environment simulator synchronized with motion capture and ball-launching systems. In the present study, we used a prototype of this system (COGNIFOOT v1—without ball-launching robots) that we installed in a turf-artificial grass playfield on which players could execute real soccer skills while facing a large screen. The whole setup is detailed below.
Playfield and support structures
The playfield size was equal to 8 × 10 × 5 m (length × width × height). Artificial-grass floor texture (PurTurf 32; Realsport ®, Rossens, Switzerland) covered the floor (see Fig. 1). Metallic structures were located around the playfield to support motion capture cameras that were placed at a height of 4.5 m (Fig. 1).
Large screen and visual environment projection
A large screen (10 × 4 m—width × height) made of a shock-absorbing tissue was located at a distance of 6 m from the ball position (Fig. 1). The visual environment was projected onto the screen using a beamer BenQ MH740 (BenQ Corporation® Taipei, Taiwan) located behind the player at a distance of 9 m to the screen and at a height of 2.9 m. The beamer was connected to a laptop (HP Elite Book; Hewlett-Packard ®, Palo Alto, CA) via the HDMI port. The generated image size was equal to 5.12 × 2.88 m (width x height). The default image background was black (same as the screen color) to ensure a constant contrast of the background over repetitions.
Real-time ball motion tracking and screen calibration
Briefly, the ball motion was tracked in real-time with 11 Optitrack Prime 17W cameras (NaturalPoint © OR, USA). Small infrared light reflective soft markers were fixed on the ball (standard diameter of 22 cm). The 3D ball position (X, Y, and Z spatial coordinates according to a reference frame centered on the initial ball position) was streamed in real time at a frequency of 360 Hz to the laptop. Ball coordinates were processed on-line using a self-written “main program” in Matlab (Mathworks Inc., Natick, MA) to compute parameters related to the CMP (see next section). The calibration of the screen was performed using four markers located at the corners of the projected image. This allowed a conversion of the target position from pixels to (Y, Z) coordinates. Postcalibration measurements guaranteed the accuracy of ball position measurements (e.g., the distance between the center of the circle displaying the ball impact position on the screen and the real ball center impact position did not exceed 1 cm). Details about this procedure can be found elsewhere (5).
Reliability of the RT measurements
The automatically computed RT were verified manually using a video-based image-by-image control procedure (see Hicheur et al. (5) for details). Nine films, recorded using a high-speed camera (CASIO EXILIM; Casio © Tokyo, Japan—sampling rate: 600 frames per second), were taken from a point of view allowing viewing simultaneously the initial ball position and the screen during a passing test. The correlation coefficient between the manually computed RT (delay between the stimulus appearance on the screen and the first movement of the ball) and automatically computed RT (COGNIFOOT v1) was equal to 0.99, guaranteeing that the computation of RT was reliable.
The properties of the visual environment (e.g., the number, location, and duration of the stimuli, inter-trial displays) were programmed using self-written Matlab routines and the Psychophysics Toolbox extensions (13–15) in Matlab. A total of 32 stimuli (one stimulus per trial) were used during a single test or training session. The target plus the distractor(s) constituted one stimulus (Fig. 1). The player faced a large screen onto which a white circular target (diameter = 0.20 m) appeared randomly (players could not anticipate the location of the upcoming target) at one of three randomly generated eccentricities (left, center, and right), and at a height of 0.10 m (so that the whole target was viewed as “lying on the floor”). The central position of the target was located directly in front of the initial ball position (distance of 6 m, 0 degree of visual angle along the eye–target longitudinal axis, perpendicular to the screen). The left and right positions were located −1.71 and +1.71 m away (−15.9° and 15.9° of visual angle) from the player. The target then immediately moved horizontally in the left or right directions, respectively. The animation (speed and direction) of the target motion were fixed differently in the pretraining/posttraining and training sessions, respectively. The target appeared without distractors in one third of the trials.
Yellow circular distractor(s) (one or two, same size as the target) appeared together with the target in two-thirds of the trials. Both distractors’ positions and movements were pseudo-randomized across trials. The stimulus duration was equal to 500 ms: the choice of such a short duration prevented players from kicking the ball before or during stimulus presentation, as tested in our previous study (5). The target position and motion direction were set so that targets appearing first in the center could move leftward or rightward while targets appearing first in the left/right parts of the screen could move rightward/leftward, respectively.
Players were asked to (a) pay attention to the target motion which will quickly disappear and (b) pass the ball as accurately and as quickly as possible toward the future position of the—nonvisible—target. In agreement with the coaches, we told players that this situation was similar to the one where they would have to pass the ball to a running teammate. Players were, thus, required to correctly anticipate the future (virtual) target location and to adjust the passing force and passing direction. They were orally instructed to find the optimal trade-off between reactiveness and passing accuracy.
Before each testing protocol, players performed six passes to become familiar with the task. They all received feedback about their performance at this stage (see Feedback Structure section).
Passing test sequence
The structure of a complete passing test with feedback is provided in Figure 1. Briefly, players were stepping in place nearby the ball which was always placed at the same initial position (6 m). To maintain a good consistency when measuring the RT, players were not allowed to execute more than one step before foot-to-ball contact after stimulus onset. A trial begun with the same sound and was followed 1 s later by the visual stimulus onset (Fig. 1). The player then executed a pass toward the (estimated) future position of the moving target. Once the ball had hit the screen, the ball was sent back to the player by two assistants who were located at the edges of the screen. A visuoauditory feedback signal was delivered at t = 8.3 to 8.5 s to all players during the familiarization trials only and to players of the AF group during all training trials (see Augmented Feedback section). The player then placed the ball at the initial position and waited for the next sound. A period of 12.3 to 12.5 s separated the end of the stimulus appearance and the sound announcing the subsequent trial so that a single trial lasted at least 13.3 s. A rest period of 25 s was included after 16 trials. Each player performed a total of 32 trials during a particular session (pretraining, training, or posttraining—see next section). The passing test typically lasted around 7 min and 30 s.
Passing test conditions
The 32 passes/trials performed by players during one session were divided into four main visuomotor conditions, which were categorized as PC-VR, PC-VL, PL-VL, and PR-VR, respectively. PC, PL, and PR denote passes toward the center, left, or right (eccentric) parts of the screen, respectively. VL and VR denote leftward or rightward visual motion directions of the target, respectively. For each visuomotor condition, the speeds of the visual target were determined with coaches during pilot tests. Hence, two target speeds were tested during both pretraining and posttraining sessions (1.20 and 1.63 m·s−1, for moderate and fast speeds, respectively) and could be combined with four types of distractors’ movements (zero distractor, one or two distractors moving in a direction similar to that of the target but at different speeds and two distractors moving in opposite directions—with one direction similar to that of the target motion), yielding a total of eight possibilities for each category of pass. The speed of distractors varied from trial to trial: it was randomly selected among speeds ranged between 100% and 144% of the moderate and fast target speeds, respectively. During the training session, the target speed also varied from trial to trial (see next section). All conditions were randomly generated so that players could not anticipate the upcoming stimulus.
We provided both visual and auditory feedback signals to inform players about their performance (through a gamification routine) while maintaining a high level of motivation throughout training sessions. The visual feedback consisted in presenting for 3 s on the screen both the ball impact position (a purple circle) and the virtual target position at the instant of impact (a blue circle showing the position of the target if it would have continued moving until impact, see Fig. 1). A sound was simultaneously provided at the beginning of the visual display. The sound (maximum duration of 2 s) was played from a sample list of 12 sound files (extracted from the game Super Mario Kart, 1992; Nintendo©) which were classified following a positive/negative reinforcement approach (16): six sounds were classified as positive and were used to reinforce accurate passes (error <30 cm), whereas six sounds were classified as negative and were used to indicate insufficient accuracy (error >30 cm). Within each category, the sound changed with 5-cm intervals (positive sound P1 was played for errors less than 5 cm, P2 was played for errors comprised between 5 and 10 cm, and so on). This gamification of training was dedicated to keeping a high level of motivation of players by informing them of both their commitment to and their progress on the task (1). All “AF group” players declared that both visual and auditory components of the feedback signal motivated them to perform better throughout training sessions.
Structure of the training protocol
All players (AF group, NF group, and CON groups) were tested before (PRE) and after (POST) the training period which lasted 17 d on average. No feedback was provided to them during these sessions (except for the six PRE “familiarization” passes described above). During the training period, all players followed a normal training program in the academy. In addition, players of the AF group and NF group performed eight COGNIFOOT-training sessions (TR01 to TR08, three sessions per week) while players of the CON group did not (and served as controls). Players of the AF and NF groups performed their COGNIFOOT training during normal training sessions. Thus, all groups (AF, NF, and CON) received a comparable overall training time during the whole intervention. Three players (one from the AF group and two from the CON group, respectively) were slightly injured during a competition game on week 2, and their remaining training/posttraining sessions were delayed by 1 wk. Players who followed the training sessions performed the posttraining tests 4 d after the last training session: this was done to exclude any fatigue effect at retesting and to favor the consolidation of potential training-related CMP improvement.
Pretraining and posttraining test sessions
The PRE session allowed us recording a baseline CMP level for all players who were re-tested in similar conditions after the training period (POST). During each of these test sessions, players perform 32 passes without receiving feedback following the conditions described above.
During each training session, players of the AF group and the NF group also performed 32 passes. However, although only four target speeds (two target speed magnitudes × two possible directions) were used in PRE and POST trials, the target speed magnitude and direction varied from trial to trial during the training session. The target speeds were randomly selected among 32 uniformly distributed speeds ranged between 100% and 144% of the moderate and fast target speeds (relatively to the PRE/POST trials’ conditions), respectively. The rationale for this was that such randomly presented target speeds (including more than half of “suprathreshold” speeds—compared with the fast speed used in PRE and POST trials) prevented players from learning only specific speeds across training sessions. Such variety in visual stimuli resulted in increased task difficulty: this was expected to facilitate CMP improvement during the POST session (see Walton et al. (1) for a discussion of how task difficulty should be increased during cognitive training). The most important difference of the training sessions between groups was that only AF group players received an augmented feedback signal (as described above) after each pass.
A total of 1728 passes were recorded during the PRE and POST sessions (27 players × 32 conditions = 864 passes for each session); 4608 passes were recorded during the training sessions (18 players × 32 conditions × 8 sessions). The trials where players shot before the appearance of the stimulus (negative RT) were excluded from the analysis. This represented a total of 6 passes (of 1728) and 42 trials (of 4608) for the PRE/POST and training sessions, respectively. As mentioned earlier, passes were divided into four visuomotor categories (passes toward the center PC-VR/PC-VL and eccentric passes PL-VL/PR-VR) based on the part of the screen where the ball was sent (center, left, or right) and the direction of the target visual motion (left or right).
Passing performance: Objective measurements
The RT (in milliseconds), the passing spatial error (PSE, in centimeters), and PS (in kilometers per hour) of players were computed automatically as described previously (see Hicheur et al. (5) for details). Briefly, RT was computed as the delay between the instant of stimulus onset and the first instant of physical ball motion. Passing spatial error was computed as the absolute distance between the ball position at impact (ball center) and the virtual position of the target at the instant of impact (e.g., the position of the target if it would have continued moving until impact). In addition, we computed a global passing performance index (GPP). The GPP index was computed as GPP = (RTp + PSEp)/2, where RTp and PSEp were expressed as percentages of minimum RT (500 ms) and PSE (20 cm) values, respectively. The GPP was higher for players with both greater reactiveness (smaller RT) and greater passing accuracy (smaller PSE). Note that we tested other ways to compute the GPP (multiplying the RT and PSE parameters in their original dimensions or testing different RT and PSE values to compute RTp and PSEp) and that this yielded similar effects.
Passing performance: Coaches and players’ judgments
Coaches were asked to judge the passing performance level of all tested players before and after the PRE and POST sessions. Players and coaches were also asked to judge potential improvements of players. Importantly, they were both told that their judgments had to be based on a passing situation on the pitch, where a particular player would have to pass the ball to a moving teammate running 5 to 10 m away from him (the closest situation to the passing test designed for the present study). Coaches were not informed about the performance of players during the passing tests when providing judgments.
They were asked to assess four aspects of the passing performance of every player. This was done through individual interviews between coaches and the same experimenter. A questionnaire had to be filled by each coach, and the role of the experimenter was to explain the assessment procedure and instructions to coaches. For every line (player) of the questionnaire table, coaches had to use three graduated five-point horizontal scales to assess, from low to high, the reactiveness (RE), the passing accuracy (PA) and the PS/power (PS, see 5 for details). Objective measurements of the passing performance were compared to coaches’ judgments. For this purpose, COGNIFOOT measurements (RT, PSE, and PS) were converted into REscore, PAscore, and PSscore using the same five-point scales used by coaches (see Hicheur et al. (5) for details). The GPPscore was computed as (PAscore + REscore)/2. The evolution of the passing performance measured by COGNIFOOT (POST minus PRE scores) was compared with coaches and players scores obtained from the questionnaires. The detailed procedure for collecting judgments was adapted from our previous study (see Hicheur et al. (5) for details) and is detailed in the Appendix (see Supplemental Material, Methods section, https://links.lww.com/MSS/B711).
We performed repeated-measure ANOVA to compare the mean performance of players (RT, PSE, PS, and GPP variables) during the PRE and POST sessions across AF group, NF group and CON group. This was done for the 1728 recorded passes and for the three groups of players (the six missing passes were replaced with the median value across players for a particular condition). Because we focused on the differential effect of the type of training on a potential performance improvement, we report in details the training (POST − PRE) × training group (AF group/NF group/CON group) interaction effect in the main article. The main effects of the pass category (PC-VR, PC-VL, PL-VL, and PR-VR), the target speed (moderate or fast), and the type of distractors’ motion (no distractor, one distractor or two distractors), as well as any significant interaction effect, are detailed in the Appendix (see Supplemental Digital Content, section Effects of training on passing performance, https://links.lww.com/MSS/B711). ANOVA were preceded by visual inspection of the normal probability plots of the residuals. In case of violations of normality, we applied Box–Cox transformation to the data (17). We also performed Levene tests to check for homogeneous variances across periods and groups’ comparisons.
The evolution of the performance gain over the eight training sessions was analyzed using repeated-measures ANOVA (N = 4608 passes, here also, the 42 missing passes were replaced with the median value across players for a particular condition of a particular training session). All ANOVA were followed up with planned contrasts.
The internal consistency of coaches’ judgments was measured using the ωh coefficient (18,19): a value of ωh equal to or above 0.7 indicates that scores are coherent across coaches (which would then validate the computation of mean coaches’ scores). We then performed ANOVA to compare the effects of the training group on the perceived performance changes across PRE and POST sessions.
Effects of Training on Passing Performance
The PRE and POST passing performance parameters are presented in Figure 2. The differential effects of the type of training on the passing performance parameters are detailed. All other statistically significant effects (main effects of training, target speed, visual distractors and category of passes, and associated interaction effects) are documented in the Appendix (see Supplemental Digital Content, section Effects of training on passing performance, https://links.lww.com/MSS/B711). Importantly, we did not observe any statistically significant difference between AF group/NF group/CON group players before training (PRE) for each of the computed performance parameters (RT/PA/GPP/PS, P > 0.05).
On average, RT were significantly shorter after training (F(1,24) = 32.0, P < 0.01, ηp2 0.57; 911 ± 103 vs 831 ± 91 ms for PRE and POST sessions, respectively). However, a statistically significant PRE/POST training x group effect (F(2,24) = 10.0, P < 0.01, ηp2 0.45), followed by planned contrasts (AF group vs NF group/CON group: t(24) = 4.46, P < 0.001) indicated that RT were significantly shorter after training only for the AF group (AF group: 943 ± 68 vs 774 ± 68 ms; NF group: 902 ± 105 vs 873 ± 92 ms; CON group: 888 ± 131 vs 846 ± 90 ms for PRE vs POST sessions, respectively; Fig. 2A). Thus, only players of the AF group significantly improved their reactiveness after training (the corresponding evolution of performance is indicated as percent in Fig. 2A).
Passing spatial error
On average, PSE (Fig. 2B) decreased by 7.43, 7.42, and 1.11 cm after training, for the AF group, NF group, and CON group, respectively. Here, we observed normality violations so PSE were normalized before ANOVA using a Box–Cox transformation (λ = 0.25; PSEn = (PSEλ − 1)/λ). Levene tests also revealed that variances were unequal across groups for the POST training period (F(2,24) = 3.93, P = 0.03). Interestingly, the Levene test was not significant when only AF group and NF group data were included (P > 0.05): the variability of PSE became significantly smaller for the AF group/NF group data compared with the CON group after training. We, therefore, excluded the CON group data from the following ANOVA. ANOVA revealed a significant effect of training on PSE (F(1,16) = 5.02, P = 0.04, ηp2 0.24), with smaller PSE after training (mean ± SD: 44.5 ± 11.4 and 37.2 ± 5.7 cm for PRE and POST sessions, respectively). No PRE/POST training–group effect was observed (P > 0.05, ηp2 0), indicating that AF group and NF group players improved their passing accuracy to a similar extent after training.
On average, PS (Fig. 2C) increased by 1.7 km·h−1 and decreased by 1.6 and 1.4 km·h−1 after training, for the AF group, NF group, and CON group, respectively. These changes were not found to be significantly affected by training or by a PRE/POST training–group effect (P > 0.05, ηp2 0.01).
Global passing performance
On average, GPP (Fig. 2D) increased by 6.93, 2.04, and 1.40 points after training for the AF group, NF group, and CON group, respectively. Here, we observed normality violations so GPP were normalized before ANOVA using a Box–Cox transformation (λ = 1.75; GPPn = (GPPλ − 1)/λ). ANOVA revealed that GPP significantly increased after training (33.3 ± 3.43 and 36.7 ± 3.59 cm for PRE and POST periods, respectively; F(1,24) = 30.0, P < 0.001, ηp2 0.55). A significant PRE/POST–group interaction effect (F(2,24) = 4.70, P = 0.019, ηp2 0.28), followed by planned contrasts (AF group vs NF group/CON group: t(24) = 9.38, P < 0.01) indicated that GPP was significantly larger after training only for the AF group (32.2 ± 2.49 and 39.2 ± 3.09 cm for PRE and POST periods, respectively). No statistically significant PRE/POST difference (P > 0.05) was observed when testing NF group and CON group. Therefore, passing performance significantly improved in the AF group only (+22%, Fig. 2D).
Performance Gain during Training
We investigated the evolution of each mean CMP parameter over training sessions (Fig. 3). ANOVA followed by polynomial contrasts (linear and quadratic) were performed on the training sessions to test for any trend describing the effect of training on each CMP parameter.
The mean RT decreased over all training sessions in the AF group while this held only for the first three sessions in the NF group (Fig. 3A). We observed a significant effect of the group (F(1,6) = 4.57, P = 0.048, ηp2 = 0.22), of the training session rank (F(7, 112) = 3.77, P < 0.01, ηp2 = 0.19), and a training session rank–group interaction effect F(7, 112) = 4.01, P < 0.001, ηp2 = 0.20). The AF group/NF group difference in RT started to diverge at T2 and increased over the following sessions, with a larger variability for NF group. Planned contrasts (AF group vs NF group, and linear/quadratic trends) revealed a different AF group/NF group linear trend across sessions (t(112) = 2.51, P = 0.0230). Performing the analysis separately on each group revealed that RT evolution over sessions followed (i) a linear and also a quadratic trend over sessions (t(56) = −2.70, P = 0.016 and t(56) = 2.14, P = 0.048, respectively) in the AF group, (ii) a quadratic trend over sessions (t(56) = 2.93, P < 0.01) in the NF group where RT of the last sessions was not different from the RT of the first training session. In contrast, AF group players improved reactiveness consistently over the training sessions (Fig. 3A).
Passing spatial error
Remarkably, although the mean PSE was comparable across groups at PRE (around 45 cm, see Fig. 2B), PSE was larger during training session 1 (T1) for both AF group (by about 2 cm) and NF group (by about 9 cm, Fig. 3B): this illustrates the higher task difficulty during training. During training, PSE was significantly smaller (and less variable) in the AF group (F(1,16) = 29.57, P < 0.001, ηp2 = 0.65; Fig. 3B) compared with NF group. PSE decreased progressively over sessions and stopped decreasing around T4 to T5 in both groups. The effect of training on PSE was close to significant (F(7, 112) = 2.03, P = 0.058, ηp2 = 0.11). We did not observe any training session rank–group interaction (P > 0.05, ηp2 = 0.04). Planned contrasts (linear and quadratic trends) revealed a significant linear decrease of PSE with training (t(112) = −2.32, P = 0.034), all other tested effects being nonsignificant (P > 0.05).
The mean PS followed a similar evolution across groups (Fig. 3C): it started to increase linearly from T2. We observed a significant effect of the training session rank (F(7, 112) = 4.73, P < 0.001, ηp2 = 0.23) but no effect of the group (P > 0.05, v) on PS. No training session rank–group interaction effect was observed (P > 0.05, ηp2 = 0.03). Planned contrasts (linear and quadratic trends) revealed a linear increase of PS across all training sessions (t(112) = 2.87, P = 0.011).
Global passing performance
The mean GPP increased from T1 to T8 in the AF group, whereas it slightly increased from T1 to T5 in the NF group (Fig. 3D). We observed a significant effect of the group (higher GPP in AF group: F(1,16) = 7.46, P = 0.015, ηp2 = 0.32) and of the training session rank (F(7, 112) = 3.00, P < 0.01, ηp2 = 0.16) but no training session rank–group interaction effect (P > 0.05, ηp2 = 0.08). Planned contrasts (AF group vs NF group) revealed no difference between these groups across sessions (P > 0.05) although the increase observed in Figure 3D seems to be more consistent in the AF group. Indeed, performing the analysis separately on each group revealed that the GPP (i) increased linearly over sessions (t(56) = 2.58, P = 0.02) in the AF group but (ii) did not follow any trend in the NF group (P > 0.05). The absence of AF group/NF group difference reported above might, therefore, be induced by a larger variability of the NF group data for almost all sessions. Interestingly, the fact that AF group players exhibited consistent improvement of the passing performance during training can explain why AF players (who had comparable performance levels before training, see previous section and Fig. 2D) outperformed NF group players after training.
Real versus Perceived Evolution of the Passing Performance in Coaches
The detailed description of coaches and players’ judgments is provided in the Appendix (Supplemental Digital Content, section Real versus perceived evolution of the passing performance, https://links.lww.com/MSS/B711).
The PRE and POST scores are presented for both COGNIFOOT and coaches (mean values across coaches) in Figure 4. Although the converted COGNIFOOT scores (left panel of the Figs. 4A to 4D) followed the same evolution as the physical parameters reported in Figure 2, a comparable increase of performance was noticeable in coaches’ judgments (right panel of Figs. 4 A toD). We, therefore, focused here on coaches’ judgments.
The RE scores were significantly higher after training (F(1,24) = 4.33, P = 0.048, ηp2 = 0.15, Fig. 4-A). No statistically significant training group effect or PRE/POST–group interaction effect were observed (P > 0.05, ηp20.09).
ANOVA revealed that PA scores were significantly higher after training (F(1,24) = 23.7, P < 0.001, ηp20.50). No significant effect of the training group (P > 0.05, ηp2 = 0) was observed. A significant PRE/POST–group interaction effect (F(2,24) = 10.5, P < 0.001, ηp2 = 0.47), followed by planned contrasts (AF group/NF group vs CON group: t(24) = 4.45, P < 0.001) indicated the presence of PA improvement only in AF group and NF group (+13.7/+9.8%, respectively, vs −2.0% for the CON group, right panel of Fig. 4B).
The PS scores were significantly higher after training (F(1,24) = 30.7, P < 0.001, ηp2 = 0.56). No significant effect of the training group (P > 0.05, ηp2 = 0) was observed. A significant PRE/POST–group interaction effect (F(2,24) = 4.13, P = 0.029, ηp2 = 0.26) followed by planned contrasts (AF group vs NF group/CON group: t(24) = 2.78, P = 0.01) indicated that PA improved significantly more in AF group compared with NF group/CON group (+9.8% vs +4.5%/+2.7%, respectively, right panel of Fig. 4C).
Global passing performance
The GPP scores were significantly higher after training (F(1,24) = 13.1, P < 0.01, ηp20.35). No significant effect of the training group (P > 0.05, ηp20.06) was observed. A significant PRE/POST–group interaction effect (F(2,24) = 4.43, P = 0.023, ηp20.27) followed by planned contrasts (AF group/NF group vs CON group: t(24) = 2.69, P = 0.013) indicated that PA improved significantly more in AF group/NF group compared with CON group (+10.0%/+5.7% vs −0.3% respectively; GPP did not change after training in CON group, right panel of Fig. 4D).
Overall, we noticed that both coaches and players perceived significant performance improvements following training (see Fig. 4 and Appendix, Supplemental Material, Figure 3sm, respectively, https://links.lww.com/MSS/B711). However, players’ evolution scores seem to be largely overestimated (including the CON group who actually did not improve at all) while the differential effect of the training group on passing performance was noticed in the coaches’ scores only (Fig. 4A1/B1/D). Because no change in passing performance was noticed in CON group players, we can conclude that coaches’ judgments are more reliable than players’ judgments.
In this study, we tested the hypothesis that augmented feedback training can improve both perceptual–cognitive and/or motor skills specific to soccer. Objective measurements of the passing performance before and after training showed that only players of the AF group significantly improved their passing accuracy and reactiveness, with a GPP improvement of 22%. In contrast, only passing accuracy improved in the NF group, whereas none of these parameters improved in the CON group. The objectively measured evolution of the passing performance was compared with the perceived evolution of the passing performance judged by coaches and players themselves. Coaches’ judgments were more reliable than players’ judgments and exhibited a training group effect comparable to the one objectively measured by COGNIFOOT. The theoretical and methodological implications of these findings for cognitive-motor training are discussed below.
Augmented-Feedback Training: Cognitive or Cognitive Motor?
As recently reviewed by Walton et al., there is a lack of direct evidence demonstrating that the enhancement of purely cognitive abilities improves sports performance. One factor explaining this limited transfer is that training one isolated (cognitive) aspect of a sport skill will not necessarily generalize to the noncognitive aspect of this skill (e.g., the motor stability) or to other skills needed for this sport. In a previous study (5), we emphasized that assessing the performance for a specific sport skill (a pass in soccer) should include both cognitive and motor aspects. In the same vein, we assumed here that cognitive-motor training (rather than a purely cognitive training) would facilitate the transfer to sporting performance. We expected that a natural effect of the repetition of visuomotor tasks would be to improve performance (1). However, improvements were mainly seen in the group receiving AF who demonstrated significantly higher and more homogeneous passing performance. In contrast, players of the NF group that accomplished the same number and type of passes but did not receive feedback only improved their passing accuracy. Thus, the additional gain of performance provided by AF does not simply reflect practice effects, as discussed in the study of Walton et al. (1). Importantly, it should be noted that the feedback provided to AF group players combined both spatial and temporal components: indeed, we displayed the virtual (future) target positions on the screen based on the instant of impact (which also depends on the reactiveness of players). Because both AF group and NF group improved their passing accuracy but only AF group players improved their reactiveness (Fig. 2), AF group players not only improved their anticipation of the future target location but they did so in a significantly shorter time compared with the NF group. Thus, AF training may improve both spatial and temporal aspects of the passing performance while a visuomotor training without feedback may be beneficial only for the spatial component, probably through practice-related motor improvements.
The analysis of the performance evolution during training sessions tends to support this possibility. Because AF cannot explain the passing accuracy improvements in the NF group during training, these can—at least partly—be explained by a practice effect (learning of the motor task). The larger performance gain in AF group players indicates an additional positive role of augmented feedback. The role of augmented-feedback was twofold: (i) it amplified the performance gain (compared to NF group) during training and (ii) it produced homogeneous improvements (spatial and temporal aspects) of the performance. The fact that AF group consistently improved both reactiveness and passing accuracy throughout training sessions (Figs. 2 and 4) suggests that performance improvement is unlikely to be the result of a strategy favoring speed over accuracy (or vice-versa) but really reflects perceptual–cognitive and motor skills’ improvement. Thus, RT/passing accuracy improvements occurred because of improved cognitive abilities (ability to estimate the speeds of the target and to compute the future location of the target based on this estimate), improved motor abilities (elaborating faster passing motor programs based on the estimation of the future target location), or some combination thereof.
The question of why performance improved less after the fourth/fifth training session in the AF group might be explained by several factors: (i) AF group players may have reached their cognitive/motor limits (a passing accuracy of ~35 cm and a RT of ~ 760 ms), (ii) the format of AF delivery (duration and/or frequency of training, and duration and/or frequency of AF delivery during training) favored acquisition of the visuomotor task but limited learning abilities of players (20), (iii) a decrease in the motivation of AF group players after a certain number of training sessions, (iv) the visuoauditory AF signal was exclusively related to the pass output: providing AF signal related to the quality of motor execution may further enhance performance improvement. A combination of these possibilities can also be considered. Besides, as described in the Methods section, we provided both visual and auditory AF to ensure a high level of motivation throughout training sessions of the AF group. Because visual feedback provided immediate information (actual ball position at impact vs virtually computed correct ball location) relative to the error of the player, it is likely that players favored visual feedback during training sessions. However, future investigations are required to test whether visual feedback alone is sufficient to drive performance enhancements reported here.
Another central question raised by our findings is related to the respective parts of purely cognitive or motor contributions to the observed performance gains. On the one side, the use of tasks of increasing difficulty should be favored during cognitive training (1). Furthermore, neurophysiological studies on perceptual learning indicate that the type of stimuli used during the training period can have a different generalization (or transfer to nonlearned stimuli) power because they can differently stimulate early or higher-order sensory areas in the brain (21). On the other side, because the incorporation of cognitive tasks into motor tasks rather than a separate training of these two aspects may generate more ecologically (or field) relevant performance gains, the first type of training might be favored in exercise and cognition researches (22). In our study, all these considerations were exploited when designing a visuomotor task inspired by a field situation naturally composed of both cognitive and motor execution requirements. In particular, the task we tested required adjusting the PS and direction based on the anticipation of the future target position. The positive role of training on the passing performance was not associated with changes in the PS (or power of the pass): this indicates the contribution of a cognitive component in the passing performance improvement. Furthermore, providing AF (informing players about the accuracy of the anticipation of the future target location) resulted in higher and more homogenous improvements of the passing accuracy (including the PR-VR category, see Appendix, Supplemental Digital Content, figure 2sm-C - PSE: PRE/POST training, Pass Category and Group, https://links.lww.com/MSS/B711) and the reactiveness (Fig. 2A) only in the AF group. This further demonstrates that cognitive abilities (the ability to anticipate the future position of a running teammate) can be efficiently trained with a task incorporating both cognitive and motor requirements.
Using indirect measurements of the passing performance during small-sided games (SSG), a recent study (23) showed that cognitive training (3D-MOT visual stimuli) improved the decisional aspect of the passing performance on the field. A combination of our approach, which allows direct measurements of the passing performance, and the one described in (23) may help to better clarify the respective roles of cognitive and motor aspects of training on the sports performance. Besides, SSG are widely used in daily soccer training but the progress produced by specific SSG can hardly be assessed (24). COGNIFOOT could also be used to monitor accurately these progresses and hence be used to design more efficient SSG.
Monitoring of cognitive-motor performance: Subjective and objective aspects
Objective measurements of various aspects of the performance are commonly used during the selection stages in elite-level youth soccer academies. However, the decision to select particular players relies largely on the subjective opinions of coaches, so far (25). This is particular true for tactical and technical aspects. The reason for this is that, together with mental aspects, cognitive-motor abilities can hardly be directly quantified in a complex and ecological valid setting. Owing to this, some authors recently argued that objective measurements are not necessarily more relevant than coach ratings (26). Would that mean that any attempt to provide objective measurements of the CMP in ecological valid situations is a vain endeavor? In the present study, we observed weak correlations when analyzing individual coach ratings and objective passing performance measurements. One could reasonably argue that this can be explained by the fact that the situation tested in our experiments does not reflect the complexity of “passing the ball to a running teammate” on the field (no physical opponents were present, for instance). However, such statement should be nuanced. First, the present results confirm our previous study where age-related changes in passing performance were rated very differently between coaches and only the mean value of all coaches’ rating resembled the one obtained with objective measures (5). Second, in the present study, the judgments of expert coaches were assessed at two different time points, and we observed a high variability across coaches when judging the evolution of passing skill performance of a single player. Therefore, there was not one representative single coach rating. Third and closely related to the previous point, only when computing mean scores across coaches (for the same individual player) resulted in performance improvements that were comparable between coaches and COGNIFOOT (Fig. 4). Forth, players’ judgments of their passing performance evolution were significantly higher (Fig. 3sm-A1/B1/C1/D) and even more variable than coach ratings (as reported in Cushion et al. (25)). Thus, providing objective measurements of the CMP seems definitely needed for accurately monitoring the baseline level and the potential progress of players after training, a season, or over successive seasons.
Methodological Considerations and Limits
The efficiency of any training program should be assessed using objective measures of the performance pretraining and posttraining. Finding objective indices of cognitive performance in team sports like soccer is a very difficult and challenging issue. One way to overcome this is to “combine physical and cognitive measures into hybrid systems, potentially a virtual reality environment” (1). Within this context, COGNIFOOT represents one example of such hybrid systems where cognitive and motor components of the passing performance are monitored in real-time. Importantly, players tested in this study were free of their movements and performed on a large playfield by delivering real passes with their feet, a situation closer to the field compared to other laboratory setups that are used to measure the CMP (see Ali (3) for a review). At the same time, a large set of finely simulated visual environments can be generated, and accurate measurements of the passing performance are realized in real time. Here, we could measure and train the capacity of players to anticipate/estimate the future position of a previously seen moving target and to adjust their pass direction and force following this estimation, a situation close to passing the ball to a running teammate. Providing such direct measurements of one component of the anticipation is, thus, of crucial interest, given that anticipation is a key factor of performance in soccer (27) which has mainly been recorded indirectly using video footages of game situations occluded at specific instants (see (28) for a review). Last but not least, we exploited here a possibility of COGNIFOOT to use real-time information about the performance to provide feedback to players: the gamification routine we implemented to adjust the auditory signal (see Methods section) likely further increased the motivation of players during training, a key feature of cognitive training (1). This factor might also be involved in the observed homogeneous performance improvements of AF group players during training, as previously discussed. Although these environmental features likely facilitate the transfer of performance to the field, several limits can be listed here. First, the simulation of a realistic passing situation in soccer should involve players and not abstract visual objects. Second, the use of a moving target displayed across a 2D environment limits the possibilities to simulate long passes. Third, players delivered passes from a static position: on the field, players most often execute passes with a ball being already in motion at the time of pass execution. All these technical limits should be overcome in the future. In addition to this methodological challenge, the challenge is to find an optimal trade-off between the need to monitor objectively the cognitive-motor performance of players and the need to take into account the more subjective opinions of coaches and players.
Augmented-feedback training is known to play a positive role for learning complex skills (10). This study provides evidence that the training of cognitive motor performance highly benefits from the use of augmented feedback, too. It seems that complex spatiotemporal patterns might only be learned with the help of objective AF, whereas the same training without AF cannot improve performance. This means that, in the future, ecological valid test and training situations should be created to objectively quantify and improve key parameters of cognitive motor performance in soccer.