Detection of season-long post-concussion motor deficits in female soccer players using a 3D goal equivalent manifold framework Authors: Xiong Zhao, Reza Rahimi, Janie Cournoyer, Allison Clouthier, Ryan B Graham Published: 2026 ABSTRACT AbstractThis study examined whether a geometry-based movement variability framework (3D Goal Equivalent Manifold, GEM) could detect persistent motor control deficits across a competitive season in female soccer players with a history of concussion, and compared its performance with the traditional 2D GEM approach. Twenty-five female varsity soccer players (11 with concussion history within 3 years, 14 without) completed bilateral bodyweight squats synchronized with a metronome (40 beats per minute) at pre-, mid-, and post-season. Whole-body kinematics were captured using a markerless motion capture system. 3D GEM analysis decomposed variability into task-relevant and task-irrelevant components, producing a 3D performance index (PI) and temporal persistence metrics (λ). Players with concussion history exhibited significantly lower 3D PI scores (M = 9.48 ± 1.60) compared to those without (M = 16.26 ± 2.45; p = 0.017), with no significant seasonal changes or interactions. No significant differences emerged for temporal persistence metrics, although the concussion group consistently showed trends toward less effective regulation in task-relevant dimensions. The 3D GEM approach effectively identified persistent motor control deficits in female soccer players with concussion histories, underscoring its potential for clinical assessments and monitoring throughout rehabilitation and return-to-play programs.Supplementary InformationThe online version contains supplementary material available at 10.1038/s41598-026-49059-0.Keywords: Concussion, Sports biomechanics, Goal equivalent manifold (GEM), Female athletes, Movement variabilitySubject terms: Health care, Medical research, Neurology, Neuroscience Introduction Sports-related concussions are prevalent in soccer, particularly among female athletes with incidence estimates ranging from 0.27 to 0.35 concussions per 1,000 athlete-exposures in matches, and even higher during competitive matches1,2. Notably, females are more likely to have prolonged recovery times and exhibit greater neurocognitive deficits post-concussion compared to males3,4. Beyond acute symptoms, growing evidence suggests that concussion can result in persistent impairments in cognitive, neuromuscular, and motor control, which may remain even after clinical symptom resolution5–7. These lingering deficits are of particular concern in high-level athletes, where performance demands require fine-tuned coordination, dynamic stability and decision making. Persistent neuromuscular control deficits can influence injury risk and performance efficiency, especially in tasks that demand rapid decision-making, balance, and inter-limb coordination8–11. Female athletes with a history of concussion often demonstrate reduced balance and altered coordination, scenarios more reflective of in-game performance7,12. Recent research has identified persistent alterations in joint coordination patterns among varsity athletes more than a year after their most recent concussion, indicating enduring deficits in inter-segmental motor control that continue to affect performance during functional tasks13. Another study also suggests that concussion may increase the risk of anterior cruciate ligament (ACL) injuries, particularly among female athletes14. These findings highlight the interconnected nature of concussion-related motor deficits and musculoskeletal injury susceptibility, reinforcing the need for comprehensive movement screening tools in this population. Traditional post-concussion assessments often rely on static balance tests (e.g., BESS) or isolated cognitive measures (e.g., reaction time), which may fail to detect subtle but functionally relevant deficits15–17. The bilateral squat is a fundamental movement pattern that challenges multiple lower extremity joints and muscle groups simultaneously, providing an integrated assessment of neuromuscular control and inter-segmental coordination18. In athletic populations, it is a universal whole-body task foundational to both training and rehabilitation. Compared to gait-based dual-task paradigms, the squat imposes greater postural and mechanical demands. When synchronized to a metronome, it introduces a temporal constraint that requires athletes to continuously integrate sensory input and motor output. Chehrehrazi et al., (2017) showed that such temporal constraints during lifting tasks can uncover meaningful differences in motor variability. Therefore, the bilateral bodyweight squat synchronized to a metronome (40 beats per minute, BPM) was selected in the present study as a sport-relevant and physically demanding task that may be especially sensitive to subtle, persistent motor control deficits in athletes with a concussion history19,20. The human neuromusculoskeletal system exhibits inherent redundancy, meaning that multiple biomechanical strategies can be employed to accomplish the same motor task. This flexibility allows for successful performance even in the presence of impairments such as those resulting from concussion, by redistributing control across joints and muscles to maintain functional output. This redundancy, also known as movement variability (MV), is reflected in the concept of equifinality: multiple optimal solutions for the same outcome, bounded by limits of adaptability and sensitivity to perturbation21,22. MV within this space can be informative; excessive variability suggests instability or exploring, while overly rigid patterns may indicate reduced adaptability23. Quantifying and interpreting this variability are central to understanding post-concussion motor control. Mathematically, the set of movement solutions that achieve the same task outcome forms a manifold. The Goal Equivalent Manifold (GEM) framework leverages this structure to evaluate MV by partitioning it into goal-relevant and goal-equivalent components24. Unlike traditional variability metrics, GEM provides insight into how movement deviations align or conflict with task demands. Recent studies have demonstrated the effectiveness of GEM analyses in capturing subtle, persistent motor control impairments that conventional assessments might overlook. For instance, Chehrehrazi et al. utilized GEM to analyze lifting variability and found significant differences in motor adaptability between healthy controls and populations with non-specific lower back pain20. These findings underscore the potential of GEM-based methodologies for identifying nuanced impairments in concussed athletes. Traditional GEM analyses have predominantly employed two-dimensional models, including variables such as displacement and velocity. However, for rhythmically constrained movements, timing precision is an equally important component of successful task execution. Therefore, this study introduces an enhanced three-dimensional GEM model by explicitly incorporating timing error as a third execution variable. Timing error quantifies deviations from the ideal cycle time, capturing subtle temporal regulation impairments that conventional 2D analyses do not detect. Comparing the results from both the traditional 2D GEM and the proposed 3D GEM will further clarify the added value and sensitivity of integrating temporal variability into the GEM framework. Thus, the objective of this study was to apply a three-dimensional GEM framework to quantify motor control deficits associated with concussion history in female soccer players. To evaluate the temporal sensitivity of these deficits, assessments were conducted at three distinct time points across the competitive season: pre-season, mid-season, and post-season. Assessing multiple time points allowed for the examination of both the persistence and evolution of neuromuscular control over time, beyond the limitations of single-time-point screening approaches. It was hypothesized that athletes with a history of concussion would demonstrate consistently lower performance indices and increased temporal variability across all testing sessions, relative to non-concussed athletes. These differences were expected to reflect persistent disruptions in neuromuscular regulation associated with prior brain injury. Methods Participants Twenty-five female varsity soccer players from the University of Ottawa were recruited, including 11 with a history of concussion (7 within one-year) and 14 without. Participants were grouped based on self-reported, medically diagnosed concussions within the past three years. Demographics (mean ± SD) for the non-concussion group were age = 20.36 ± 2.17 years, height = 1.64 ± 0.07 m, weight = 61.15 ± 5.38 kg, and BMI = 22.72 ± 1.81 kg/m². For the concussion group age = 20.45 ± 1.81 years, height = 1.69 ± 0.05 m, weight = 66.06 ± 9.94 kg, and BMI = 23.13 ± 2.89 kg/m². Inclusion criteria were (1) age ≥ 18; (2) full medical clearance for athletic participation; and (3) ability to complete 40 repetitions of a bilateral bodyweight squat to a metronome, used as a measure of coordination and neuromuscular control. This study was approved by the Research Ethics Board at University of Ottawa (file #: H-07-24-10628). All participants provided written informed consent prior to participation and were informed of their right to withdraw at any time without consequence. All data were anonymized prior to analysis to ensure confidentiality. All methods were performed in accordance with the relevant guidelines and regulations, including the Declaration of Helsinki. Protocol Testing was conducted at three time points: pre-season (week 1), mid-season (week 9), and post-season (week 15). Upon arrival, participants provided informed consent, completed an injury history questionnaire, and had their height and weight measured (pre-season only). After familiarization, participants performed 35 bodyweight squats to a metronome (Fig. 1 bottom) at 40 BPM. Participants began each squat from an upright position with fully extended hips and knees. A full repetition involved descending until the thighs reached at least parallel to the ground while maintaining heel contact, followed by a return to the upright position. Participants were instructed to squat down and stand up in sync with the metronome beat to the best of their ability (Fig. 1). No additional constraints were imposed to allow natural movement variability. Whole-body kinematics were captured using an 8-camera markerless motion capture system (Vicon, Oxford, UK) at 60 Hz and analyzed using Theia3D (Theia Markerless Inc., Kingston, ON, Canada). During the same data collection session, participants completed a series of additional movement screening tasks, a single bilateral trial of T-Balance and Y-Balance, one bilateral L-Hop trial, three repetitions of bilateral countermovement jumps, three broad jumps, and two reaction time tasks presented in randomized order. These assessments were part of a broader testing battery but are beyond the scope of the present analysis. The full testing protocol, excluding consent form and questionnaire, was repeated at mid- and post-season. Demonstration of the bilateral body-weight squat to a metronome: (a) start upright with hips and knees fully extended; (b) descent toward thighs at least parallel to the floor; (c) bottom position; (d) ascent from the bottom; (e) return to full extension. Data analysis After data collection, videos were processed using Theia3D (Theia Markerless Inc., Kingston, ON, Canada) and analyzed in Visual3D (C-Motion Inc., Germantown, MD, USA) to extract each participant’s center of mass (COM) kinematics. The kinematic data were exported to Matlab R2019b (The MathWorks Inc., MA, USA) for processing. Custom scripts were used to apply a fourth-order zero-lag Butterworth filter with a 6 Hz cutoff frequency. To ensure consistency and that they reached steady-state motion, only the final 30 repetitions were retained for analysis. Each repetition was segmented based on the COM vertical trajectory. For each rep, displacement and velocity of the COM were computed using Euclidean distance and first-derivative of displacement, respectively. Timing error was calculated as the deviation (in seconds) between the actual start and end of each repetition and the ideal timing based on the 40-bpm metronome cue (3s per rep: 1.5 s down and 1.5 s up). The motion of the body’s COM in the sagittal plane was used to represent the primary control variable for this movement19,25. With the body-level variables being the displacement (D) and the velocity (V) of the COM, the goal is maintaining constant time (T) through all trials, and the goal is defined as Eq. (1): where is the displacement of the COM for cycle n, and is the linear velocity of the COM for that cycle. This relationship was set equal to a target movement time (i.e., 3 s), leading to a scalar goal function Eq. (2) that applies to each observed pair However, because the task (squatting to a metronome) demands precise temporal control, a timing error term that explicitly captures deviations from the target timing was introduced, acting as a dimension of execution variability. In other words, the timing error, , was modeled as a third execution (body-level) variable that directly parameterizes the global timing constraint, thereby linking body-level control to the task-level goal mathematically. Timing error represents the signed deviation from the metronome for each repetition. Although it appears explicitly in the task equation, it functions as one dimension of the execution space together with displacement and velocity. The 3D GEM framework therefore evaluates how variability across these execution variables is distributed relative to the task constraint surface, rather than simply quantifying the magnitude of timing error itself. This expands the manifold to represent spatio-temporal solutions rather than purely spatial (2D GEM) solutions and addresses a limitation of the 2D GEM, in which any cycle-time deviation is implicitly treated as goal-relevant error even when D and V covary in a coordinated way. The revised goal function, defined for each trial-specific set of values , is given in Eq. (3): This formulation implies that for a given execution triple , the task goal is perfectly achieved when the measured ratio is exactly equal to the target time . A nonzero thus reflects the extent to which a trial deviates temporally from the ideal execution, and by integrating this component into the GEM framework, both kinematic and temporal aspects of the performance can be assessed. To compare MV between participants in the proposed analysis, D and V were normalized to each participant’s standard deviations across cycles26. The two-dimensional approach was extended by incorporating timing error () as a third dimension which was also normalized by each participant’s standard deviation across cycles to reflect the temporal demands of the task. Any combination of normalized values () that satisfies this equation lies on the goal-equivalent manifold (GEM) (see Fig. 2, black dash line). A perfect task execution, , in this case should equal to three seconds. To define the task-consistent operating point (), we computed the point on the GEM that minimized the Euclidean distance to the mean of execution vector, defined by the average displacement, velocity, and timing error, across all repetitions. Specifically, this operating point was obtained by solving a constrained optimization problem where the objective was to minimize Eq. (4): Visualization of the two-dimensional goal equivalent manifold (GEM) analysis of the squat. Individual dots represent individual repetitions. The diagonal dash line stands for the GEM, and are unit vectors for variations along the GEM and deviations perpendicular to the GEM. subject to the nonlinear constraint , ensuring the operating point satisfied the task timing requirement. The resulting operating point () was then normalized using the standard deviation of displacement and velocity across trials to align with the normalized 3D execution space used in subsequent GEM decomposition. Variability relative to the operating point is computed as Eq. (5): To capture the structure of motor variability, the deviation vector for each cycle was given by ]. This decomposition was obtained by first calculating the gradient (or Jacobian) of the goal function at the operating point () (Eq. 6): After normalizing, , the null space of defines the directions in the 3D execution space along which variability does not affect the task outcome (i.e., the goal-equivalent manifold or GEM). Let denote a 3 × 2 matrix, where each column corresponds to an orthonormal tangent vector to the GEM manifold. The variability vector was then decomposed into goal-equivalent variability (), i.e. variability along the GEM, which does not affect task performance and goal-relevant variability (), i.e. variability perpendicular to the GEM, which directly influences performance. Mathematically, this is expressed as Eq. (7): To quantify performance, the overall Performance Index (PI) is defined as the ratio of the standard deviation of the goal-equivalent variability to that of the goal-relevant variability Eq. (8): A greater 3D performance index () indicates that a larger proportion of the variability is directed along the GEM (i.e., “good” variability that does not negatively impact task performance), reflecting superior task performance19,20. An example of the 3D GEM plot is shown in Fig. 3. In addition to the primary analysis using standard deviation scaling, a sensitivity analysis was conducted for 3D PI calculation using robust z-score normalization based on the median and median absolute deviation (MAD). Raincloud plot of 3D PI for Non-Concussion (blue) and Concussion (orange) groups across Pre, Mid, and Post season stages. Each half‐violin displays the full distribution of individual PI values (dots), while the dashed horizontal lines denote group mean. To further assess the temporal persistence of trial-to-trial variability, the lag-1 autocorrelation coefficients (, , and ) for each of the goal-equivalent directions (, ) and the goal-relevant direction () were computed using Pearson correlation between its value in trial n and trial n + 1 across the 30 repetitions. This reflects the degree of temporal dependence in the respective variability dimensions. Higher λ values indicate greater persistence in the corresponding dimension, whereas values near zero suggest more random or adaptive behaviour across repetitions19,25. 2D GEM analyses were conducted to provide a direct comparison with our proposed 3D GEM framework. The calculations followed the same procedure described above, except that timing error, , was excluded. As a result, only two execution variables, displacement and velocity, were used to compute the goal-equivalent variability , goal-relevant variability , and their corresponding lag-1 autocorrelation metrics ( and ). The 2D performance index () was also calculated as defined in Eq. (8) with only and . To evaluate the added variance of incorporating timing error into the GEM framework, the total variability in both the [ ] and execution spaces was computed. For each time point, the variance of was calculated, and its proportion relative to the total 3D execution variance was computed using the Eq. (9): This analysis provided an estimate of how much temporal variability contributed to the overall execution variability in the metronome-paced squat task. Statistical analysis All statistical analyses were performed using IBM SPSS Statistics, Version 25.0 (IBM Corp., Armonk, NY, USA). Prior to modeling, the distribution of each continuous outcome variable was assessed using Shapiro-Wilk tests. A generalized linear model with generalized estimating equations (GEE) for repeated measures was used to examine the effects of stage, concussion history, and their interaction on 2D/3D performance index. A Gamma distribution with a log link was used to model the right-skewed performance index, with an M-dependent (1) correlation structure and robust covariance estimation. GEE estimation was used to account for within-subject correlations and allowed the inclusion of participants with partially missing data27. Stage was modeled as a categorical predictor with pre-season as the reference. Parameter estimates from the generalized linear model with GEE estimation were used to assess changes over time, with pre-season set as the reference category. Estimated coefficients (B) reflect the mean difference (M) in the outcome at each stage relative to pre-season, along with standard errors (SE) and associated Wald χ² tests. Where applicable, pairwise comparisons used Sequential Bonferroni adjustment. Significance was set at p < 0.05. To assess the impact of missing data, a sensitivity analysis was conducted by only including participants who completed all pre-, mid- and post-season sessions. Group mean difference was compared at each stage between completers and all data available before rerunning the generalized linear model with GEE. To assess whether participant evaluations on 2D PI and 3D PI were consistent, monotonic (Spearman’s ρ) and rank (Kendall’s τ) correlations between 2D PI and 3D PI were computed at each time-point. Nonparametric measures and exact p-values were reported when distributions were skewed. To evaluate the influence of extreme observations, Spearman’s ρ was repeated after 5% symmetric winsorization of both 2D and 3D PIs, and a leave-one-out analysis was conducted (recomputing ρ after removing each participant in turn). Potential outliers in PI were identified by group using the Tukey quartile rule. For a simple between-group snapshot, 2D PI and 3D PI were compared between non-concussion and concussion groups using the Mann-Whitney U test (two-sided α = 0.05). Results In the pre-season (week 1), 23 participants completed the in-lab session, with one defender (concussion) and one midfielder (non-concussion) absent (non-concussion: concussion = 13:10). In the mid-season (week 9), 24 participants completed the session, with one non-concussed midfielder missing (13:11). In the post-season (week 15), 20 participants completed the session, with five players absent: two forwards (one concussion, one non-concussion), one non-concussed midfielder, and two defenders (one concussion, one non-concussion) (11:10). When the model was limited to participants with complete data, no mean difference was found, and the significance remains for both 2D and 3D analyses. Independent-samples Welch t-tests showed no significant differences between groups in age (p = 0.82), height (p = 0.28), weight (p = 0.31), or BMI (p = 0.74). Results were consistent when restricting the analysis to complete cases (age p = 0.72, height p = 0.92, weight p = 0.49, BMI p = 0.51). At pre-season, timing error contributed approximately 14.74% (0.35 / 2.35) of the total variability. This contribution was lower at mid-season at 6.34% (0.14 / 2.14) and rose again post-season to 15.64% (0.37 / 2.37). On average across the season, timing error explained 12.24% of the total movement variability. Analysis of raw timing error showed no significant group differences across stages (p = 0.857). Group means for 2D and 3D performance index and their temporal persistence metrics across all three time points are summarized in Tables 1 and 2, respectively. Table 1Summary statistics of 2D performance index and temporal persistence metrics (, ) by group and season stage.StageGroupPI2D PreConcussion1.08 ± 0.160.56 ± 0.23-0.40 ± 0.20Non-concussion1.15 ± 0.310.41 ± 0.20-0.41 ± 0.16MidConcussion1.07 ± 0.220.46 ± 0.19-0.44 ± 0.12Non-concussion1.28 ± 0.650.36 ± 0.23-0.43 ± 0.16PostConcussion1.11 ± 0.210.41 ± 0.30-0.42 ± 0.19Non-concussion1.27 ± 0.540.40 ± 0.24-0.22 ± 0.28Note: Values are presented as mean ± standard deviation; PI2D = 2D Performance Index; , = lag-1 autocorrelation of task-irrelevant components; = temporal persistence within the task-relevant space in 2D GEM. Summary statistics of 2D performance index and temporal persistence metrics (, ) by group and season stage. Note: Values are presented as mean ± standard deviation; PI2D = 2D Performance Index; , = lag-1 autocorrelation of task-irrelevant components; = temporal persistence within the task-relevant space in 2D GEM. Table 2Summary statistics of 3D performance index and temporal persistence metrics (, , ) by group and season stage.StageGroupPI3D* PreConcussion7.17 ± 4.250.14 ± 0.270.00 ± 0.34-0.24 ± 0.25Non-concussion16.40 ± 16.880.05 ± 0.300.08 ± 0.34-0.16 ± 0.29MidConcussion9.28 ± 5.31-0.01 ± 0.190.00 ± 0.37-0.27 ± 0.24Non-concussion13.59 ± 9.66-0.05 ± 0.220.04 ± 0.37-0.26 ± 0.21PostConcussion12.44 ± 12.820.05 ± 0.250.00 ± 0.42-0.25 ± 0.31Non-concussion19.60 ± 16.700.03 ± 0.250.17 ± 0.33-0.13 ± 0.28Note: Values are presented as mean ± standard deviation; * = statistical significance was found; PI3D = 3D Performance Index; , = lag-1 autocorrelation of task-irrelevant components; = temporal persistence within the task-relevant space. Summary statistics of 3D performance index and temporal persistence metrics (, , ) by group and season stage. Note: Values are presented as mean ± standard deviation; * = statistical significance was found; PI3D = 3D Performance Index; , = lag-1 autocorrelation of task-irrelevant components; = temporal persistence within the task-relevant space. 2D PI values ranged from 1.07 to 1.28 across all groups and time points, with generally lower values in the group with concussion history at all time points, though not statistically significant. The GEE results indicated no significant main effect of concussion history on 2D PI, χ² (1) = 1.57, p = 0.21, nor for stage, χ² (1) = 1.54, p = 0.46, or their interaction, χ² (1) = 1.42, p = 0.49. For , values ranged from 0.36 to 0.56, and for , values remained consistently negative (-0.44 to -0.22) across all time points. Similarly, no significant effects were observed for (concussion: χ²(1) = 2.11, p = 0.15; stage: χ² (1) = 1.30, p = 0.52; stage × concussion interaction: χ² (1) = 0.59, p = 0.75) or (concussion: χ²(1) = 1.89, p = 0.17; stage: χ²(1) = 2.83, p = 0.24; stage × concussion interaction: χ² (1) = 3.09, p = 0.21). At each stage of the season, players with a history of concussion exhibited lower 3D PI scores compared to those without a concussion history (Table 2). The largest discrepancy was seen at the pre-season stage (7.17 vs. 16.40), and although the concussion group improved over time and the non-concussion group dipped at mid-season, the non-concussion group consistently maintained significantly higher 3D PI values (Fig. 4). When components of the 3D PI, goal-equivalent (“good”) and goal-relevant (“bad”) variability, were examined separately. Goal-equivalent variability was similar between groups across the season (Pre: 1.44 vs. 1.46; Mid: 1.45 vs 1.45; Post: 1.44 vs. 1.46), whereas goal-relevant variability was higher in the concussion group (Pre: 0.20 vs. 0.28; Mid: 0.25 vs. 0.27; Post: 0.17 vs. 0.33). Detailed values are provided in Supplementary material 1. Players in the both groups tended to show random or adaptive behaviour in task-relevant space (, ). However, the concussion group exhibited consistently lower values in . 3D GEM visualization of one participant. Each point shows a repetition in normalized COM displacement, velocity, and timing error space. Colors also indicate timing error for better visualization in 3D space, with darker ones lagging the beat and lighter ones leading. The grey plane represents the GEM plane; and represent variability along and perpendicular to the GEM, respectively. The main effect of concussion history on 3D PI was significant, χ² (1) = 5.74, p = 0.017, with players without a history of concussion (M = 16.26, SE = 2.59) exhibiting higher 3D PI values than those with a concussion history (M = 9.34, SE = 1.57). Neither the main effect of stage, χ² (2) = 2.79, p = 0.248, nor the Stage × Concussion interaction, χ² (2) = 1.69, p = 0.428, were statistically significant. When execution variables were normalized using robust z-score scaling, the 3D PI showed a similar directional trend but did not reach statistical significance (χ² = 2.80, p = 0.094) (Supplementary material 2). Notably, normalization using standard deviation scaling remains a commonly used approach in GEM analyses to create dimensionless execution variables and was therefore retained as the primary scaling method in the present study. No significant effects of Stage, Concussion or Stage × Concussion interaction were observed for and . Similarly, no significant effects emerged for in 3D analysis, which reflects trial-to-trial regulation in task-relevant performance space (Fig. 5). Estimated marginal means indicated that players with concussion history consistently exhibited lower values across all stages (pre: M = -0.24, SE = 0.08; mid: M = -0.27, SE = 0.07; post: M = -0.25, SE = 0.10) than players without concussion history (pre: M = -0.16, SE = 0.08; mid: M = -0.26, SE = 0.06; post: M = -0.13, SE = 0.08). Bar plots of , , and values for Non-Concussion (blue) and Concussion (orange) groups at Pre, Mid, and Post seasons. Each bar shows the group mean and error bars represent ± 1 SD. As shown in Table 3, across pre-, mid-, and post-season sessions, 2D and 3D PIs were consistently and positively associated (overall Spearman ρ ≈ 0.64–0.68; Kendall τ ≈ 0.45–0.48). Within-group associations were positive at each stage, with the strongest correlation in the non-concussion group at post (ρ = 0.827). Robustness checks (5% winsorization and leave-one-out) yielded nearly identical estimates and narrow ranges, indicating the findings were not driven by single observations; occasional 3D PI Tukey outliers were flagged (pre: one non-concussion; post: one non-concussion) and one potentially influential case at mid (non-concussion), but inferences were unchanged. Single-timepoint rank-sum tests detected no significant non-concussion vs. concussion differences for either 2D or 3D PI at any stage (pre: 3D p = 0.369, 2D p = 0.556; mid: 3D p = 0.271, 2D p = 0.817; post: 3D p = 0.224, 2D p = 0.879). Table 3Summary of concordance between 2D and 3D PIs at pre-, mid-, and post-season, with within-group correlations and robustness checks.StageOverall ρ (p)Overall τ (p)NC ρ (p)C ρ (p)Winsorized ρ (p)Leave-one-out ρ rangePre0.676 (0.001)0.478 (0.001)0.659 (0.017)0.721 (0.024)0.676 (0.001)0.630–0.729Mid0.640 (0.001)0.471 (0.001)0.676 (0.014)0.673 (0.028)0.640 (0.001)0.598–0.748Post0.657 (0.002)0.453 (0.005)0.827 (0.003)0.650 (0.067)0.657 (0.002)0.625–0.718Note: ρ = Spearman rank-order correlation; τ = Kendall’s tau. NC = non-concussion; C = concussion. Values are correlations with two-sided exact p-values in parentheses; complete-case analyses only. Winsorized ρ was computed after 5% symmetric winsorization of both performance indices. The leave-one-out range is the minimum-maximum Spearman ρ obtained when iteratively removing each participant once. Summary of concordance between 2D and 3D PIs at pre-, mid-, and post-season, with within-group correlations and robustness checks. Note: ρ = Spearman rank-order correlation; τ = Kendall’s tau. NC = non-concussion; C = concussion. Values are correlations with two-sided exact p-values in parentheses; complete-case analyses only. Winsorized ρ was computed after 5% symmetric winsorization of both performance indices. The leave-one-out range is the minimum-maximum Spearman ρ obtained when iteratively removing each participant once. Discussion This study examined the impact of concussion history on motor control deficits in female soccer players using GEM analysis in both 2D and 3D. We hypothesized that players with a history of concussion would exhibit persistently lower motor control performance indices (PI) and altered temporal persistence in motor variability across a competitive season compared to players without concussion histories. Additionally, we anticipated these deficits would remain stable over the season, highlighting lasting neuromuscular impairment despite medical clearance. By treating timing error as a body-level variable tied to the global constraint, the 3D GEM separates temporal regulation from spatial coordination, classifies coordinated spatio-temporal trade-offs as goal-equivalent, and isolates only goal-relevant error, thereby increasing sensitivity to post-concussion deficits in a rhythm-constrained task. Examination of the 3D PI components showed that goal-equivalent (“good”) variability was comparable between groups across the season, whereas goal-relevant (“bad”) variability was higher in the concussion-history group. Thus, the lower PI may primarily arise from increased bad variability. The inclusion of timing error as an additional execution dimension in the GEM framework was shown to meaningfully contribute to the total variability captured in the task. Specifically, timing error accounted, on average, for approximately 12% of the overall execution variability across the competitive season. Although no universal threshold exists for interpreting additional variance, literature suggests that modest increases in explained variance (5–15%) can provide meaningful insights28. Given the inherent rhythmic demands of this task, explicitly modeling temporal variability enhances the practical applicability of the GEM approach, capturing nuances of neuromuscular control that a spatial-only (2D) analysis would not detect. Both 2D and 3D GEM analyses revealed that female soccer players with a history of concussion consistently demonstrated lower PI across the season. However, only the 3D GEM model showed statistically significant group differences. This suggests that including temporal variability may provide additional information about task-relevant motor control during cyclic movements. Further, as shown in Tables 2 and 3, PIs exhibited consistent, moderate-strong positive concordance overall and within both groups, indicating that the spatio-temporal index (3D PI) largely preserves the athlete ranking captured by the spatial-only index (2D PI) while extending it to include timing regulation. Robustness checks (identical 5% winsorized ρ and narrow leave-one-out ranges) show that these associations are not driven by single observations; the few flagged 3D PI outliers (non-concussion at pre and post) did not change the inferences. Coupled with non-significant single-timepoint group comparisons, this pattern supports the utility of the 3D GEM, by explicitly incorporating timing error, to detect season-long, concussion-related decrements in performance regulation. Sensitivity analyses further indicated that the 3D PI was likely influenced by the normalization method used to scale the execution variables. Since GEM decompositions are based on geometric projections within the execution space, the relative scaling of displacement, velocity, and timing error can alter the distribution of variability and the resulting PI values. These findings highlight the importance of consistent variable normalization when applying multi-dimensional GEM models and suggest that interpretation of 3D PI should consider the potential influence of execution-space scaling. The finding aligns with previous literature highlighting persistent deficits in neuromuscular function, including reduced coordination, impaired balance, and altered dynamic stability, well beyond clinical recovery and symptom resolution3,7,9. Studies have reported persistent motor deficits in athletes post-concussion, emphasizing the need for extended assessment periods to capture subtle, ongoing impairments. Interestingly, no significant seasonal variations or interaction effects between concussion history and stage of the season were observed, suggesting that these motor control impairments are stable rather than transient or context-dependent (e.g., training load). Such consistency underscores the long-lasting impact of concussions on neuromuscular control systems, potentially heightening vulnerability to recurrent injuries or reduced performance quality over prolonged periods12. Chen et al. similarly reported stable inter-joint coordination deficits in concussed individuals, suggesting these impairments are robust and enduring29. These stable motor control deficits might not be evident through traditional clinical assessments, emphasizing the importance of employing sensitive analytical methods like GEM to detect subtle yet clinically relevant impairments. Temporal persistence metrics (λ values) exhibited non-significant yet consistent trends toward less effective trial-to-trial motor regulation in players with concussion histories. Previous studies suggest such subtle trends might be linked to impaired neural processing and diminished cognitive-motor integration post-concussion5,10,11. This diminished motor regulation capacity could reflect an ongoing disruption in neural mechanisms responsible for motor planning and adaptation, further highlighting the complexity of complete functional recovery after concussion. For example, Chen et al. reported increased variability in hip-knee coordination during gait, hypothesizing a reduced cognitive processing capacity as a contributing factor29. This is consistent with broader evidence indicating persistent deficits in attention, cognitive flexibility, and executive function post-concussion30. Additionally, Fino et al. demonstrated increased lower extremity injury rates post-concussion, suggesting impaired motor planning and adaptive responses could significantly elevate subsequent injury risks10. Lynall et al. found similar heightened risks of musculoskeletal injury in athletes who recently recovered from concussion, emphasizing the clinical importance of identifying persistent motor control impairments early and accurately11. Given GEM’s sensitivity in detecting subtle deficits, it presents a valuable tool for clinicians to objectively assess and tailor rehabilitation strategies. Integrating GEM analysis into routine clinical practice can provide quantitative insights that help refine individualized rehabilitation protocols, monitor athlete progress more effectively, and support informed return-to-play decisions, ultimately enhancing athlete safety and performance resilience31,32. Several studies have recommended targeted neuromuscular interventions aimed explicitly at improving coordination, stability, and adaptive motor control post-concussion. For instance, targeted neuromuscular training has been shown to effectively reduce injury risk and enhance motor performance in post-concussion recovery phases33,34. However, specific rehabilitation programs addressing altered inter-segmental coordination post-concussion remain underexplored, highlighting an important area for future investigation. Ippersiel et al., speculated similar altered coordination patterns might represent adaptive strategies to compromised proprioception in low back pain populations, a hypothesis applicable to concussion populations as well34. By incorporating timing error as a third dimension into the execution space, this modified GEM framework advances beyond conventional two-dimensional applications. It enables a more holistic analysis of motor control by capturing spatial variability and temporal precision, critical in tasks demanding rhythmic coordination. Unlike the Uncontrolled Manifold (UCM) framework, which evaluates how athletes manage variability to achieve consistent outcomes without requiring an explicit goal function, the GEM approach is especially valuable in sports assessments because it incorporates a clearly defined goal function. This allows for detailed analysis of how athletes regulate performance by coordinating both body-level and task-level variables to achieve task-specific goals24. Collectively, the combination of a repetitive task under temporal constraint and a refined 3D GEM method offers a valid framework for evaluating post-concussion motor control. It facilitates the identification of persistent deficits that may not be captured through conventional assessments, supporting a more nuanced understanding of functional recovery. Limitations This study has its limitations. First, although the generalized linear model with GEE is well-suited to handle missing data, the relatively small sample size constrained the statistical power and limited the generalizability of the findings. While consistent trends were observed, a larger sample would increase confidence in the robustness of observed group differences. Second, participants with a concussion history varied in terms of the time elapsed since their most recent concussion. The heterogeneity in injury recency likely introduced variability in the extent of motor control impairment, potentially masking time-dependent recovery patterns. Third, the bodyweight squat was performed at a fixed pace synchronized to a 40 BPM metronome. Although this moderate speed was selected based on a pilot study and logistical constraints related to athlete availability, using a single preset tempo may not have accurately captured each athlete’s natural movement rhythm or their maximal motor control capacity. Future research should continue exploring GEM’s utility across diverse athletic populations and tasks, including assessments at varying movement speeds to better reflect individual control capacities. Investigations should also compare acute versus chronic stages post-concussion and examine how motor control deficits evolve over time and relate to subsequent injury risk. Longitudinal research employing advanced biomechanical and electrophysiological methods could provide deeper insight into the mechanisms underlying persistent motor impairments post-concussion. Further studies should also investigate the effectiveness of neuromuscular training interventions specifically designed to improve motor coordination deficits identified through 3D GEM analysis, thus offering direct, evidence-based recommendations for clinical rehabilitation practices. Conclusion This study demonstrated the utility of 3D GEM analysis in identifying persistent motor control deficits in female soccer players with concussion histories. Clinically, these findings carry substantial implications. They emphasize the necessity for ongoing and detailed neuromuscular monitoring beyond traditional symptom management Rehabilitation specialists could use GEM analysis to evaluate the effectiveness of neuromuscular training interventions over time, adapting programs in real time to the athlete’s performance trajectory. This approach supports a shift toward function-based, personalized rehabilitation models, ensuring that athletes regain not only symptom-free status, but also optimal biomechanical resilience needed for sport-specific demands and highlights the clinical value of advanced analytical tools such as GEM. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Supplementary Material 2 Acknowledgements The authors would like to sincerely thank all the athletes from the University of Ottawa’s women’s soccer team for their time, effort, and commitment throughout the study. We are also grateful to the coaching staff for their invaluable support, coordination, and facilitation of data collection during the competitive season. This study would not have been possible without their collaboration. Author contributions X.Z. conceived the study, implemented the 3D GEM pipeline, performed all analyses, and drafted the manuscript. R.R. collaborated on the mathematical formulation, algorithm optimization and code verification. J.C. facilitated the recruitment and interpretation of the results. R.B.G. and A.C. supervised the project and provided funding, conceptual guidance, support with data analysis, interpretation of results, visualization, and critical review and editing. All authors contributed to manuscript revision and approved the final version. Funding Xiong Zhao and Ryan B. Graham were funded by University Research Chair (URC) from the University of Ottawa and Natural Sciences and Engineering Research Council (NSERC) Allison Clouthier was funded by Natural Sciences and Engineering Research Council (NSERC), and Janie Cournoyer and Reza Rahimi did not receive any funding for this study. Data availability Data are available upon reasonable request. Declarations Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Supplementary Materials Supplementary Material 1 Supplementary Material 2 Data Availability Statement Data are available upon reasonable request.