Reliable electrocortical dynamics of target-directed pass-kicks

Eleven healthy participants (3 female/8 male, mean age: 27.42 ± 3.68), who were not engaged in sports on a regular basis, participated in this study. Only right-dominant participants were recruited to avoid hemispheric bias (Marcori et al. 2020). The laterality of the participants was determined based on Lateral Preference Inventory (Coren and Porac 1978). None of them had previous orthopaedic injuries or neurological diseases. Their activity level was assessed with the Marx Activity Scale (Marx et al. 2001) and an additional questionnaire, which indicated activity only at a recreational level. All subjects had normal or corrected vision at the time of the experiment. They were informed about the purpose and procedures of the study and gave written consent prior to participation. The ethical committee of the affiliated university approved the conduction of this study in accordance with the Declaration of Helsinki.

The present study adopted a simplified kicking task in order to ensure a stable, comparable movement pattern among novices. The target was the rectangular frontal surface (10 × 15 cm) of a wooden box fixated on the floor. In contrast to larger goals on the field, a much smaller and precise target was chosen in order to eliminate the vertical dimension of accuracy and have a concrete description of precision (Hunter et al. 2018). A standard soccer ball (FIFA size 5) was placed three metres apart from and perpendicular to the target. Participants placed their left foot next to and their right foot behind the ball with a light external rotation and were instructed to perform a pass-kick with the inside of their foot (Zago et al. 2014). They performed 10 trial-kicks to familiarise with the task and during these kicks, they were allowed to fine-tune the position of their feet to assign the most comfortable location, which was subsequently marked on the floor for the stance leg to assure a standard position for all trials. Finally, in this position, self-initiated pass-kicks were performed with the dominant (right) leg towards the target as accurately as possible. The trials in which the participants were able to hit the rectangular surface with the ball were counted as hits. In total, 90 trials were performed within six blocks of 15 trials. An overview of the experimental setting is presented in Fig. 1.

The three dimensional biomechanics of the kicking and stance leg were recorded using a wearable inertial measurement unit (IMU) system (myoMOTION, Noraxon, USA) at a sampling rate of 200 Hz. Seven IMU sensors were placed on the pelvis (sacral surface), thighs (laterally on the lower quadrant), shanks (anteromedially on tibial surface) and feet (on metatarsal surface dorsally and proximally to the ankle; Berner et al. 2020). The recorded biomechanical data was digitised and exported using myoRESEARCH Software (version 3.14, Noraxon, USA). The maximum acceleration of the kicking foot in the x-axis corresponding to ball contact (Lees et al. 2010) was extracted for each kick using MATLAB (version R2020b, The Math Works, USA). The minimum, maximum and median values among all 90 kicks were calculated for each participant. To ensure that participants performed pass-kicks with comparable biomechanics described in other studies, the movement range of hip flexion, knee flexion and foot external rotation was computed (Kellis and Katis 2007; Zago et al. 2014). Additionally, the accuracy performance was video-recorded using a webcam (Logitech Brio, Switzerland) synchronised with the aforementioned software. The number of hits was determined and the accuracy rate was calculated in percentage as the proportion of hits to the total number of kicks (Lepschy et al. 2018). In order to verify a comparable kicking pattern, speed and accuracy performance between two sessions, the test–retest reliability of the movement range for the aforementioned joints, minimum, maximum and median acceleration values of the kicking foot and accuracy rate were computed based on intraclass correlation coefficients (ICC; Henriksen et al. 2004).

Cortical activity was continuously recorded throughout the experiment using 65 active electrodes (actiCap, Brain Products, Germany) and a mobile amplifier (LiveAmp64, Brain Products, Germany). The electrodes were positioned in accordance with the international 10–20 system, with AFz being the ground and FCz being the reference electrode (Pivik et al. 1993). A 3D acceleration sensor (Brain Products, Germany) directly connected to the mobile amplifier was placed on the posterior of the lateral malleolus to detect onset of the kicks. The impedance was reduced to 25 kΩ and the EEG signal was recorded using BrainVision Recorder (Brain Products, Germany) at a sampling rate of 500 Hz.

The recorded EEG data were processed using MATLAB (version R2020b, The Math Works, USA) and the EEGLAB toolbox (version 14.1.2b, Delorme and Makeig 2004). Sinusoidal line noise was removed using the Cleanline plugin (Mullen 2012) and the signal was band-pass filtered between 3 and 30 Hz. Automatic detection of noisy channels was based on a deviation criterion, and the channels, whose robust z-score of standard deviation was more than 5, were removed. To avoid bias in the repeated measures design due to the unequal number of channels per session, mssing channels were interpolated before running ICA (Bidgeley-Shamlo et al. 2015). The data were then re-referenced to a common average and downsampled to 256 Hz.

The preprocessed data were epoched based on kick-onset. The kick-onset was detected based on the 3D acceleration data as the point where a statistically significant increase was observed in the transverse plane, namely the x-axis (Lees et al. 2010). The acceleration signal was rectified and smoothed with a Gaussian-weighted moving average filter at a window length of 1000 (Mendi et al. 2013). Subsequently, abrupt changes in the resulting signal corresponding to kick-onset were determined on the basis of signal mean using the ischange function in Matlab. In order to focus on peri-event cortical modulations such as movement planning, execution and monitoring, an epoch time window of 3000 ms before and after kick-onset was chosen with the following reasons: (i) Preparatory cortical processes have been shown to emerge two seconds before the movement onset (Shibasaki and Hallett 2006), (ii) The approximate duration of the kicks in our data was two seconds (0–1000 ms: backswing and swing, 1000–2000 ms: follow-through), (iii) Shortening of epochs after wavelet transformation. Upon the visual inspection of epochs, those containing non-stereotypical artefacts were rejected. Baseline correction was performed from − 2500 to − 2000 ms (Groppe et al. 2009) and the data were decomposed into maximally independent sources of cortical activity using an adaptive mixture independent component analysis (AMICA; Palmer et al. 2012). To avoid the effect of rank deficiency due to spherical interpolation, the dimension of the analysis of the principal components was reduced relative to the number of channels interpolated (Bidgeley-Shamlo et al. 2015). The spatial source of decomposed independent components (IC) was estimated using a standardised four-shell spherical head model (BESA, Germany) implemented in the DIPFIT plugin (Oostenveld and Oostendorp 2002). Correspondingly, brain components were labelled based on their source, activity and residual variance (≤ 15%, Onton and Makeig 2006). The brain components detected in the first session and subsequently in both sessions were clustered using the k-means algorithm in two different steps. To avoid circular inference in the subsequent statistical analysis, clustering was only based on dipole locations (Kriegeskorte et al. 2009). ICs located more than three SDs apart from cluster centroids were set as outliers. The number of clusters was specified using the Silhouette, Davies-Bouldin and Calinski-Harabasz optimisation algorithms based on the mean distance of ICs to cluster centroids (Miyakoshi et al. 2020). Reproducible clusters representing at least 50% of the sample in both clustering steps and including ICs of both sessions in at least 50% of the sample in the second clustering step (Peterson and Ferris 2018; Solis-Escalante et al. 2019) were assumed to be prominent cortical areas involved in kicking and considered in reliability analysis.

For the reproduced clusters, participants contributing with ICs of both sessions were retained. For each participant and IC, the ERSP matrix was extracted for session I and II and individually for the frequency ranges of interest, namely theta (4–7 Hz), alpha-1 (8–10 Hz), alpha-2 (11–13 Hz) and beta-1 (14–20 Hz) using the integrated study function in EEGLAB. Due to possible contamination caused by muscular activity at frequencies higher than 20 Hz, the beta-2 range was sidelined (Paluch et al. 2017). The corresponding ERSP matrices were accessed in MATLAB and for participants with multiple ICs in a session, the ERSP values were averaged on a pixel basis in order to transform multiple ICs into one representative array for a single session (Delorme and Makeig 2004). This procedure ended up in a 50 × 200 matrix with each cell representing a single pixel in the ERSP map with x axis corresponding to epoch time points and y axis to frequency. Subsequently, the ICC estimates (r values) and the corresponding p-values were computed for each pixel in order to investigate test–retest reliability (Shrout and Fleiss 1979). With this approach, a resultant reliability map was achieved at which the x- and y-axes indicated time and frequency, respectively, and the consistency in ERSP patterns between session I and I was demonstrated (the visualisation of ICC estimation protocol is presented in online source 1). Finally, for the interpretation of cortical dynamics in a systematic manner, the r values were averaged for four frequency bands of interest (theta, alpha-1, alpha-2 and beta-1) and for each 250 ms in a 4 × 20 matrix.

All statistical analyses were performed in MATLAB (Version 2020b, The Math Works, USA). Initially, the normality of the acceleration data was assessed using MATLAB’s swft function (Gardner-O’Kearny 2021) and the median value was adopted to indicate the average peak acceleration for each participant due to skewed distribution. The relative test–retest reliability of behavioural and cortical data was investigated using ICC function (Salarian 2021). The ICC estimates and their 95% confidence intervals (CI) were computed with the following formula of a two-way mixed effects model, based on single ratings and absolute agreement (McGraw and Wong 1996; MSBS = mean square between subjects; MSE = mean square for error; k = number of measurements; n = number of subjects; MSBM = mean square between measurements).

An ICC value is conventionally a positive value between 0 and 1 (negative estimates in some cases), with values between 0.5 and 0.75 indicating moderate reliability, 0.75 and 0.9 good reliability and > 0.9 excellent reliability (Koo and Li 2016). In the current study, these values were adopted to rate reliability and the negative ICC scores, as well as the pixels with a non-significant p-value, were treated as zeros (Zhang et al. 2011; Zhang et al. 2017). Systematic differences resulting from random variation between two time points were inspected using paired t-test.

The analysis of the biomechanical data demonstrated that participants performed pass-kicks with backswing and swing phases at both sessions. The hip flexion, knee flexion and foot external rotation components are shown in Fig. 2. The movement range of hip flexion, knee flexion and foot external rotation showed moderate reliability from session I to II (ICC = 0.65, CI = 0.02–0.92; ICC = 0.62, CI = 0.06–0.89; ICC = 0.74, CI = 0.20–0.94 respectively). There were no systematic differences in any of the mentioned movement ranges between both sessions.

The minimum and the median values of the peak acceleration showed moderate reliability (ICC = 0.68, CI = 0.16–0.91; ICC = 0.57, CI =  − 0.02 to 0.87 respectively), whereas the maximum value showed poor reliability (ICC = 0.25, CI =  − 0.39 to 0.74). No systematic differences were found between sessions.

The accuracy rate showed moderate reliability (ICC = 0.56, CI =  − 0.002 to 0.86) from session I (M = 63.94, SD = 12.04) to II (M = 66.19, SD = 12.45, for indivudual accuracy rates please see online source 2).

The optimum number of clusters suggested by the three optimisation algorithms (Silhouette, Davies-Bouldin and Calinski-Harabasz) was five for both clustering steps. The cortical data recorded at the first session revealed a right parieto-occipital cluster (11 participants, 18 ICs, Fig. 3A), a right fronto-parietal cluster (5 participants, 10 ICs, Fig. 3B), a left parieto-occipital cluster (6 participants, 14 ICs, Fig. 3C), a left frontal cluster (5 participants, 8 ICs, Fig. 3D) and a mid-frontal cluster (8 participants, 15 ICs, Fig. 3E). The second clustering step with ICs of both sessions revealed a right parieto-occipital cluster (11 participants, 43 ICs, Fig. 3F), a right fronto-parietal cluster (10 participants ICs, 24 ICs, Fig. 3G), a left parieto-occipital cluster (6 participants, 20 ICs, Fig. 3H), a left fronto-parietal cluster (10 participants, 27 ICs, Fig. 3I) and a mid-frontal cluster (7 participants, 25 ICs, Fig. 3J). The number of components per participant showed moderate to excellent reliability from session one to two (ICC = 0.87, CI = 0.61–0.96). The right parieto-occipital (Fig. 4) and the mid-frontal cluster (Fig. 5) represented at least 50% of the participants at both time points and included components from both sessions for at least 50% of the participants (100% and 64% respectively), and were considered for reliability analysis.

ERSPs of the right parieto-occipital cluster (Fig. 6) demonstrated a strong desynchronization in the alpha band upon kick-onset at both sessions. The desynchronization emerged in the alpha-2 band in the swing phase, radiated to alpha-1 and slightly to beta-1 bands following ball contact, becoming more evident prior to ball contact between 1000 and 2000 ms in the alpha-2 band. The ICC estimates showed moderate to excellent reliability for these frequency bands after kick-onset with higher scores after 1000 ms overlapping with the stronger desynchronisation in the follow-through phase. The averaged ICC estimates are presented in Fig. 7.

ERSPs of the mid-frontal cluster (Fig. 8) revealed a theta synchronization starting prior to ball-contact and continuing in the swing phase. At both sessions, the synchronization was distinctive after kick-onset and in the second half of the swing phase towards ball contact. Furthermore, a desynchronization could also be seen in the alpha band starting in the follow-through phase and becoming stronger peri-kick-end. The ICC estimates showed moderate to good reliability for the theta synchronization observed at kick-onset and in the swing phase before ball contact until 1000 ms. For the alpha desynchronization seen after ball contact, the ICC estimates demonstrated moderate to excellent reliability. The averaged ICC estimates are presented in Fig. 9.