Topological data analysis of postural sway: characterizing balance control during childhood development Authors: Elaheh Azadian, Mahdi Majlesi Published: 2026 ABSTRACT AbstractBackgroundPostural control matures through gradual refinement of sway organization, yet traditional CoP measures capture only variability and not the underlying geometric structure. Topological Data Analysis (TDA) offers a complementary approach by quantifying the persistence and complexity of sway dynamics. This study applied TDA to investigate developmental and task-related differences in CoP topology from childhood to adulthood.MethodsA total of 460 healthy participants (ages 7–12 years and young adults) completed eight standing balance tasks that manipulated visual input, base of support (BoS), and cognitive load. For each 30-s CoP trial, Vietoris–Rips persistence diagrams were computed, and four H1-based TDA metrics were extracted: H1_Wasserstein distance, H1_TotalPersistence, Persistent Entropy, and H1_NumLongLoops. Mixed-design ANOVAs evaluated the effects of Age, Gender, Vision, BoS, and Cognitive Dual-Tasking (CDT) on each metric.ResultsAll TDA metrics exhibited strong age effects (all p < 0.001), with children showing greater persistence magnitudes and more long-lived loops than adults, reflecting less stable and more oscillatory sway organization. Vision significantly modulated H1_Wasserstein and Total Persistence (p < 0.001), with eye closure producing pronounced increases. Single-leg stance robustly amplified all TDA measures (p < 0.001). CDT induced moderate but significant alterations in several metrics, whereas Persistent Entropy remained insensitive to visual deprivation (p = 0.168). Gender differences were generally small but interacted with task difficulty for selected outcomes.ConclusionTDA revealed systematic developmental transitions in the topology of CoP dynamics, shifting from rich multi-loop structures in childhood to topologically simpler and more stable configurations in adulthood. These findings support TDA as a sensitive and physiologically meaningful approach for evaluating balance maturation and may aid in detecting atypical postural control patterns in clinical or developmental contexts.Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1186/s12984-026-02078-4.Keywords: Topological data analysis, Postural balance, Center of pressure, Visual deprivation, Cognitive load Introduction Balance control during upright stance depends on the coordinated integration of visual, vestibular, and somatosensory inputs that guide corrective motor responses [1]. Throughout childhood and adolescence, postural control undergoes profound maturation as sensorimotor integration improves and sway patterns become more refined. With increasing age, children show reduced and less variable Center of Pressure (CoP) excursions, and by adolescence their balance performance largely approximates adult levels [2]. Younger children, however, use different postural strategies, display larger and more irregular sway, and have greater difficulty coping with challenging conditions [3]. The ability to efficiently integrate visual information improves notably around 7–10 years of age, whereas children aged 4–6 often struggle to combine visual and proprioceptive cues [4]. Similarly, attentional capacity and dual-task performance continue to mature across childhood, with cognitive loading destabilizing younger children more than adults [5, 6]. These developmental differences highlight the need for sensitive quantitative tools to characterize the maturation of postural control. Conventional quiet-standing assessments primarily rely on linear CoP metrics such as sway amplitude, velocity, or RMS displacement [7]. While these indices quantify sway magnitude, they provide limited insight into the temporal structure or complexity of postural adjustments [8, 9]. To overcome these limitations, nonlinear approaches—particularly entropy-based measures—have been introduced to capture the regularity and dynamical complexity of CoP trajectories [10]. Higher entropy values reflect more complex and less predictable sway, features generally associated with a flexible and adaptable postural control system [6, 11]. Healthy young adults typically exhibit an optimal level of sway complexity, whereas aging and various neurological disorders are linked to reduced complexity [12]. However, entropy measures summarize overall signal irregularity and may not fully represent underlying dynamical patterns. Notably, in some studies, visual deprivation increased sway magnitude without systematically altering sample entropy, suggesting that linear and nonlinear measures probe distinct aspects of postural control [11, 13]. Topological Data Analysis (TDA) has recently emerged as a powerful approach to quantify complex dynamics by characterizing the shape of data [14]. TDA methods are grounded in algebraic topology and are robust to noise and choice of metric [15], making them well-suited for physiological time-series analysis. In TDA, several topological constructs are central to understanding the data, including connected components, loops, and voids. Connected components refer to subsets of points in the data that are closely connected to each other without any separation, forming coherent regions. Loops represent cyclic or recurrent structures that emerge in the data, reflecting oscillatory or repeating patterns over time. Voids correspond to regions lacking significant topological structure. These features enable TDA to capture the geometric organization of data across multiple scales, providing insight into structural patterns that are not readily accessible through traditional analyses [16]. Nonlinear approaches such as entropy-based measures and recurrence analysis have been widely applied to characterize postural sway dynamics. These methods primarily quantify temporal irregularity, predictability, or recurrence properties of the signal, offering important information about system complexity. However, they do not explicitly describe how data points are geometrically organized in space or whether the signal exhibits stable structural patterns across scales. In contrast, TDA provides a complementary perspective by quantifying the persistence of geometric features within the data, allowing identification of robust structures such as cyclic trajectories or spatial clustering patterns. This distinction is particularly relevant for postural control, where the organization of sway trajectories—not only their temporal variability—reflects underlying sensorimotor strategies. By capturing these persistent geometric features, TDA offers additional insight into the spatial structure and stability of postural sway that may not be detected by conventional temporal complexity metrics [12, 15]. The primary tool in TDA is persistent homology, which analyzes data (e.g., a signal or point cloud) across multiple spatial scales to extract topological features such as connected components, loops, and voids [14]. The result is a persistence diagram (PD), a set of points each representing a topological feature with coordinates corresponding to the “birth” and “death” of that feature as the scale grows [17]. Features that persist over a wide range of scales (i.e. points far from the diagonal in the PD) are presumed to represent meaningful structure in the data, whereas short-lived features (close to the diagonal) are likely noise [18]. By quantifying these patterns, TDA can detect recurrent or cyclic structures in time-series data that other methods might miss. For example, in a classic dynamic system with limit-cycle behavior, persistent homology would identify a robust loop corresponding to the recurring cycle [19]. TDA has been successfully applied in various biomechanics and neurophysiology contexts. In gait analysis, persistent homology-based features have helped detect subtle abnormalities in patients with neurological diseases that were not captured by standard gait metrics [20, 21]. For instance, Bois et al. (2022) used TDA to compare gait time-series of multiple sclerosis patients and healthy controls, showing that topological features can distinguish subtle gait changes related to disease severity [22]. Similarly, Tong et al. (2021) combined TDA with machine learning for Parkinson’s disease classification, using Persistent Entropy – an information-theoretic summary of persistence diagrams – to encode gait dynamics for input into a classifier [17]. These studies demonstrated that TDA-derived metrics (such as persistence diagrams, barcodes, and their summaries) can serve as sensitive features reflecting underlying motor control differences. Despite its promise, TDA has not yet been widely applied to postural control development. Most previous research on children’s balance used linear sway measures and basic nonlinear indices (entropy, fractal measures), with few examining the topological structure of CoP trajectories across age [7, 23]. We hypothesize that TDA can reveal novel insights into how children’s postural sway patterns evolve with maturation and under different task conditions. In particular, recurrent sway patterns or oscillatory corrections might manifest as topological loops in the state-space of the CoP dynamics. A more unstable or exploratory sway (expected in younger children) may produce multiple H1_NumLongLoops in the persistence diagram, whereas a highly stable posture (expected in adults) might yield fewer or shorter persistence features. This study examines static postural control in children and young adults using TDA. Four TDA-derived metrics from the CoP time series were analyzed: H1_Wasserstein distance, H1_TotalPersistence (TP), Persistent Entropy, and H1_NumLongLoops. These measures capture complementary aspects of sway topology. The Wasserstein distance quantifies the overall magnitude of loop features in the persistence diagram [24], while TP reflects the cumulative lifetimes of these features. Persistent Entropy characterizes the dispersion and complexity of the persistence distribution [17], and H1_NumLongLoops indexes the number of long-lived, prominent loops representing meaningful cyclic sway patterns. By comparing these metrics across age groups, genders, and task conditions (vision, BoS, CDT), we aim to identify developmental and task-dependent changes in the topological structure of postural sway. We expect TDA to complement classical CoP analyses—for example, through age-related reductions in persistence or by revealing metrics that are less sensitive to sensory perturbations, as noted previously for sample entropy. Demonstrating the clinical utility of these topological markers may ultimately support TDA-based tools for identifying developmental delays or guiding targeted balance interventions. Methods Participants This cross-sectional study included 460 healthy volunteers comprising 420 children (ages 7–12 years) and 40 young adults. The child cohort was stratified into six age groups (7, 8, 9, 10, 11, and 12 years), with each age group consisting of 35 girls and 35 boys to ensure balanced representation across developmental stages. The young adult reference group included 20 females and 20 males (mean age = 24.3 ± 1.5 years). The study design primarily focuses on age-related changes within childhood, while the adult group serves as a comparative reference rather than a continuous extension of the developmental sample. All participants were free from neurological, musculoskeletal, vestibular, or uncorrected visual impairments that could influence postural control, based on self-report for adults and parental report for children. Additional exclusion criteria included a history of lower-limb injury within the past six months or the use of medications affecting balance. The sample size was primarily determined by the availability of eligible participants and is substantially larger than in previous developmental balance studies, providing adequate statistical power for repeated-measures analyses. In the present study, the dominant leg was identified as the leg the participant would use to kick a ball as far as possible [25]. The study protocol was approved by the Ethics Committee of Islamic Azad University (IR.IAU.H.REC.1400.002) and conducted in accordance with the Declaration of Helsinki. Written informed consent was obtained from all adult participants and from parents or legal guardians of minors, with children providing verbal assent prior to participation. Protocol and task conditions Participants completed eight standing postural tasks designed to systematically manipulate sensory and cognitive demands. The tasks followed a 2 × 2 × 2 factorial structure comprising: visual condition (eyes open vs. eyes closed), BoS (feet together vs. single-leg stance), and cognitive dual task (CDT) (no secondary task vs. dual task). This design was adapted from established paradigms used in dual-task postural control research [26]. In the eyes open condition, participants fixated on a black target placed at eye level, 3 m in front of them; in the eyes closed condition, they maintained the same stance with visual input removed. BoS was manipulated by instructing individuals to stand either with feet together (a stable bipedal posture) or on their dominant leg (a more challenging unipedal stance). During single-leg stance trials, the non-dominant leg was flexed with the foot elevated off the ground. For the CDT condition, participants performed a digit-span working memory task (Wechsler subtest), consistent with prior studies [2]. A sequence of random digits was read aloud, and participants were asked to repeat the digits in the correct order while simultaneously maintaining their posture. The number of digits correctly recalled within a 20-second window was documented to verify active engagement in the CDT. Each participant performed three trials of each of the eight conditions (total = 24 trials). Trial order was randomized to minimize sequence and learning effects. Each trial consisted of 30 s of quiet standing, performed barefoot to standardize somatosensory input. Participants were instructed to remain as still as possible without consciously stiffening or restricting natural sway. Short, self-paced rest intervals were provided between trials to reduce fatigue. Throughout testing, an experimenter remained close to the participant for safety monitoring, but no physical assistance was provided unless a loss of balance occurred. Data acquisition and preprocessing Postural sway was quantified using CoP trajectories recorded with a Kistler force platform (Type 9286BA, 40 × 60 cm; Kistler Instrument AG, Winterthur, Switzerland) at a sampling frequency of 1000 Hz. Ground reaction forces were transformed into anterior–posterior (AP) and mediolateral (ML) CoP coordinates using standard biomechanical procedures. To ensure signal stability and remove transient artifacts, the first and last 2 s of each trial were discarded prior to analysis. The CoP signals were then low-pass filtered at 10 Hz using a fourth-order zero-phase Butterworth filter to attenuate measurement noise and high-frequency components [27]. Subsequently, the AP and ML time series were mean-centered. For computational efficiency in TDA, the CoP trajectories were downsampled by a factor of 4 prior to further processing. This reduction preserves the overall geometric structure of the trajectories while substantially reducing computational cost. No spatial normalization of CoP coordinates was applied. The trajectories were analyzed in their original measurement space to retain absolute geometric characteristics of postural sway. Although anthropometric factors and base of support may influence CoP geometry, this approach allows condition- and group-related differences to be captured as they manifest in the observed trajectories. For data quality assessment, conventional linear sway metrics (including path length and mean velocity) were computed. In addition, sample entropy (m = 2, r = 0.2 × SD) was calculated to verify signal consistency. These measures were used solely for quality control and were not included in the primary analyses [7]. Topological data analysis pipeline An overview of the TDA pipeline is presented in Fig. 1. The present analysis does not involve state-space reconstruction or delay embedding and therefore does not aim to infer the full nonlinear dynamics of the postural control system. Because the aim was to characterize the geometric structure of sway rather than reconstruct a full dynamical attractor, no delay embedding or phase-space reconstruction was applied [17, 28]. Instead, TDA is applied directly to the two-dimensional CoP trajectories to characterize their geometric and topological structure in the measurement space. Topological Data Analysis (TDA) pipeline applied to Center-of-Pressure (CoP) signals. A Example of the static single-leg stance task performed with eyes closed under a cognitive dual-task condition on a force platform. B Preprocessed CoP time series in the anteroposterior (AP, black line) and mediolateral (ML, red line) directions over a 30-s interval. C Planar CoP point-cloud representation in the AP–ML plane, used directly as input for persistent homology without delay embedding. D Persistence diagram obtained from the Vietoris–Rips filtration, displaying H₀ (connected components; green points) and H₁ (loop features; red points), where points farther from the diagonal indicate features with greater persistence. E H₁ barcode illustrating the birth–death lifetimes of loop features across filtration values, where longer horizontal bars represent more persistent topological structures. Colors indicate different topological feature types (H₀ vs. H₁) rather than magnitude scaling. TDA metrics, including H1_Wasserstein distance, H1_TotalPersistence, Persistent Entropy, and H1_NumLongLoops, are derived from the persistence diagram and barcode representations Prior to TDA, CoP signals were preprocessed by removing the first and last 2 s of each trial to eliminate transient effects. The signals were then low-pass filtered at 10 Hz using a fourth-order zero-phase Butterworth filter and subsequently mean-centered in both AP and ML directions. To reduce computational burden while preserving the global geometric structure of the trajectories, the CoP time series were downsampled by a factor of 4 (i.e., every fourth sample was retained). The resulting two-dimensional trajectory for each trial was represented as a sequence of points in the AP–ML plane: where x(t) denotes the CoP position at time t, and CoPAP(t) and CoPML(t) represent the coordinates in the AP and ML directions, respectively. Each trajectory was treated as a point cloud for persistent homology analysis [29, 30]. Persistent homology was computed using a Vietoris–Rips filtration implemented via the Ripser library, with homology calculated up to dimension 1 (H₁) [31]. Only H₁ features (loop structures) were retained for subsequent analyses. Pairwise Euclidean distances in the AP–ML plane were used to construct the filtration. No explicit filtration threshold was imposed; the filtration scale was determined automatically based on the data. From each persistence diagram, topological features were characterized by their birth (bi) and death (di) radii [32], and the lifetime of each feature was defined as: where Li represents the persistence of the i-th feature. Let n denote the number of H₁ features in the persistence diagram, indexed by i = 1, …, n. Four quantitative TDA metrics were extracted: Wasserstein-based metric: Wasserstein-based metric: This expression corresponds to the L2 norm of feature lifetimes and provides a computationally efficient summary of overall loop persistence. It can be interpreted as a lifetime-based measure of the prominence of H₁ features relative to the diagonal of the persistence diagram. (2)H1_TotalPersistence: H1_TotalPersistence: which represents the cumulative persistence of all loop features [33, 34]. (3)Persistent Entropy: Persistent Entropy: where pi denotes the normalized contribution of the ith feature. Entropy is then computed as: which quantifies the distributional complexity of feature lifetimes [35]. (4)H1_NumLongLoops: H1_NumLongLoops: The number of long-lived loop features was defined as the number of H₁ features whose persistence exceeded 20% of the maximum persistence observed within the corresponding persistence diagram: where Li denotes the persistence (lifetime) of the i-th H₁ feature. This relative threshold was selected to emphasize dominant topological structures while reducing the influence of short-lived features that are more likely to reflect local geometric fluctuations or noise. In addition, persistence images were computed to characterize the spatial distribution of topological features using the Persim implementation of PersistenceImager with a pixel size of 0.05. The image domain was automatically fitted to each persistence diagram using the library’s default fitting procedure. Image mean and standard deviation were subsequently extracted as summary descriptors. All topological computations were performed in Python using the Ripser and Persim libraries together with custom analysis scripts [31, 36]. To improve reproducibility, all preprocessing and topological analyses were implemented using custom Python scripts. The same preprocessing sequence (trimming, filtering, mean-centering, and downsampling), Vietoris–Rips filtration procedure, and feature-extraction pipeline were applied identically to all participants and experimental conditions. Topological summaries were derived directly from the two-dimensional AP–ML CoP trajectories rather than from delay-embedded state-space reconstructions. Statistical analysis We examined the effects of Age, Gender, and Task conditions on each TDA metric using a mixed-design ANOVA. Age group (7, 8, 9, 10, 11, 12, adult) and gender (female, male) were treated as between-subject factors, whereas vision (eyes open vs. eyes closed), BoS (feet together vs. single-leg stance), and CDT (single vs. dual-task) served as within-subject (repeated-measures) factors, resulting in a 7 × 2 × 2 × 2 × 2 factorial design for each outcome. Prior to conducting the analyses, all dependent variables were inspected for normality using Shapiro–Wilk tests and Q–Q plots. Outliers were screened using standardized residuals, and log-transformations were evaluated when deviations from normality were detected. Because analyses on raw and transformed data yielded qualitatively similar results, only the raw-data results are reported for interpretability. The omnibus ANOVA tested for main effects and interactions among all factors. Given the large number of potential higher-order interactions, we focused a priori on interpretable interactions involving Age (to characterize developmental differences) and those involving Vision or BoS, which were expected to exert the strongest task-related modulations of postural control. Mauchly’s test of sphericity was applied to repeated-measures factors, and Greenhouse–Geisser corrections were used when sphericity was violated. Significant ANOVA effects were followed by Bonferroni-adjusted post-hoc comparisons. Pairwise contrasts between age groups were examined at each level of task difficulty, and simple effects of Vision and BoS were assessed within individual age groups. We additionally evaluated whether Gender differences varied by age by testing simple effects of Gender within each age group. Effect sizes for ANOVA results are reported as partial eta-squared (ηp²). Although a multivariate approach (MANOVA) was considered due to correlations among the TDA metrics, univariate ANOVAs are reported for clarity. While the TDA metrics are theoretically designed to capture different aspects of sway dynamics, the observed correlations indicate partial overlap among them; therefore, each metric is interpreted as reflecting related but not entirely independent characteristics of postural control. Given the correlated nature of these outcomes and our strong a priori hypotheses, we interpreted patterns of effects and effect sizes rather than applying strict corrections across all four metrics. To facilitate interpretation relative to established biomechanical measures, exploratory Pearson correlation analyses were conducted between TDA-derived metrics and conventional CoP variables. These results are reported in the Supplementary Material. All statistical analyses were performed using IBM SPSS v21 and Python (NumPy/SciPy). Statistical significance was set at α = 0.05. Results Table 1 summarizes the demographic characteristics of the participants stratified by age group and gender. Table 1Demographic characteristics of the participants, means (SD)Age(year) N Height (m)Weight (kg)BMIFemaleMaleFemaleMaleFemaleMale7701.23 (0.05)1.22 (0.05)23.75 (3.57)22.85 (2.70)15.47 (1.78)15.26 (1.08)8701.27 (0.06)1.29 (0.06)24.79 (4.81)26.64 (4.44)15.13 (2.28)15.78 (1.82)9701.33 (0.07)1.34 (0.06)30.39 (6.86)29.44 (6.10)16.79 (2.51)16.25 (2.43)10701.41 (0.04)1.40 (0.06)32.49 (5.80)33.42 (7.03)16.24 (2.50)16.98 (2.91)11701.46 (0.04)1.47 (0.06)37.77 (5.19)38.32 (7.46)17.59 (2.42)17.60 (2.55)12701.52 (0.06)1.48 (0.05)41.59 (9.55)39.12 (5.98)17.85 (3.33)17.77 (2.01)Adult401.62 (0.06)1.82 (0.07)52.32 (6.73)84.26 (10.3)19.94 (2.74)24.67 (2.70)Each child group: 35 girls and 35 boys; adults: 20 females and 20 males; Values are presented as mean (SD)BMI Body Mass Index Demographic characteristics of the participants, means (SD) Each child group: 35 girls and 35 boys; adults: 20 females and 20 males; Values are presented as mean (SD) BMI Body Mass Index The results of the mixed-design ANOVA for all TDA-derived metrics are summarized in Table 2. Interaction effects across visual, stability, and cognitive conditions are illustrated in Fig. 2. Table 2Results of the mixed-design ANOVA for the four TDA-derived metrics across age, gender, visual condition, base of support, and cognitive dual-task, including all two-way interactions [F-value (p-value)]FactorsH1_WassersteinH1_TotalPersistencePersistent EntropyH1_NumLongLoopsAge10.51 (0.001)*17.85 (0.001)*28.14 (0.001)*9.63 (0.001)*Gender17.26 (0.001)*28.06 (0.001)*0.50 (0.480)2.43 (0.120)Age× Gender3.08 (0.006)*3.13 (0.005)0.83 (0.550)1.46 (0.190)Visual639.64 (0.001)*993.03 (0.001)*1.91 (0.168)0.17 (0.681)Visual× Age14.62 (0.001)*25.31 (0.001)*0.251 (0.959)1.42 (0.233)Visual× Gender0.09 (0.759)1.85 (0.174)1.20 (0.275)0.91 (0.487)BoS192.69 (0.001)*912.38 (0.001)*1084.21 (0.001)*700.83 (0.001)*BoS× Age5.90 (0.001)*17.25 (0.001)*8.28 (0.001)*11.42 (0.001)*BoS× Gender2.53 (0.113)1.42 (0.234)36.69 (0.001)*5.89 (0.014) *Cognitive load40.54 (0.001)*33.96 (0.001)*16.98 (0.001)*2.82 (0.094)Cognitive load× Age2.52 (0.021)*2.19 (0.043)*3.55 (0.002)*2.63 (0.016)*Cognitive load× Gender11.18 (0.001)*20.93 (0.000)*12.20 (0.001)*1.71 (0.191)Visual× BoS468.39 (0.001)*727.73 (0.000)*27.89 (0.000)*16.26 (0.000)*Visual× Cognitive load1.47 (0.226)4.66 (0.031)*1.30 (0.255)0.011 (0.916)BoS× Cognitive load1.31 (0.253)14.62 (0.000)*35.09 (0.000)*7.20 (0.008)*Values are F (p)BoS Base of support, Cognitive load Cognitive dual task* indicates statistical significance at p < 0.05 Results of the mixed-design ANOVA for the four TDA-derived metrics across age, gender, visual condition, base of support, and cognitive dual-task, including all two-way interactions [F-value (p-value)] Values are F (p) BoS Base of support, Cognitive load Cognitive dual task * indicates statistical significance at p < 0.05 Interaction effects of visual, stability, and cognitive conditions across age and gender on TDA features. Interaction plots illustrating how visual condition (EO, EC), stability level (S1, S2), and cognitive load (C1, C2) interact with age and gender to influence topological data analysis (TDA) metrics of postural control. Panels present H1_Wasserstein distance, H1_TotalPersistence, Persistent Entropy, and H1_NumLongLoops. Error bars represent ± 1 SE. Abbreviations: EO = eyes open; EC = eyes closed; S1 = stable stance (feet together); S2 = unstable stance (single-leg stance); C1 = single-task; C2 = dual-task. A visual gap between the oldest child group (12 years) and the adult group is introduced to reflect the discontinuity between developmental stages rather than a continuous age progression H1_Wasserstein Significant main effects emerged for visual condition (F(1,425) = 639.64, p < 0.001, η²=0.60), BoS (F(1,425) = 192.68, p < 0.001, η²=0.31), and CDT (F(1,425) = 40.54, p < 0.001, η²=0.09). Eye closure and single-leg stance increased H1_Wasserstein distance by approximately 43% and 24%, respectively, whereas cognitive loading produced a 9% reduction. Age (F(6,425) = 10.51, p < 0.001, η²=0.13) and gender (F(1,425) = 17.26, p < 0.001, η²=0.04) also showed significant main effects, with all child groups displaying higher values than adults and boys exceeding girls by about 10%. The age × gender interaction was significant (F(6,425) = 3.08, p = 0.006, η²=0.042), indicating that gender differences were limited to ages 7–8. A significant age × visual interaction (F(6,425) = 14.62, p < 0.001, η²=0.17) showed that removing vision increased H1_Wasserstein distance in all child groups (all p < 0.0001) but not in adults (p = 0.724). The gender × visual interaction was not significant. The age × BoS interaction (F(6,425) = 5.90, p < 0.001, η²=0.08) indicated that single-leg stance elevated values across all ages except adults (p = 0.737). The age × CDT interaction (F(6,425) = 2.52, p = 0.021, η²=0.034) showed that cognitive load reduced values in all groups except ages 7 and 12. Finally, CDT effects remained significant within each gender, reducing H1_Wasserstein distance in girls (p = 0.034) and boys (p < 0.001). H1_Total Persistence (TP) Significant main effects were observed for visual condition (F(1,425) = 993.04, p < 0.001, η²=0.70), BoS (F(1,425) = 912.38, p < 0.001, η²=0.68), and CDT (F(1,425) = 33.96, p < 0.001, η²=0.07). Eye closure increased TP by approximately 46%, single-leg stance by 49%, whereas CDT produced a modest reduction of about 6%. Age (F(6,425) = 17.85, p < 0.001, η²=0.20) and gender (F(1,425) = 28.06, p < 0.001, η²=0.06) also yielded significant main effects, with adults showing the lowest TP, children aged 7–8 the highest, and boys exceeding girls by roughly 11%. The age × gender interaction was significant (F(6,425) = 3.13, p = 0.005, η²=0.04), indicating that gender differences were confined to ages 7–8. A robust age × visual interaction (F(6,425) = 25.31, p < 0.001, η²=0.26) demonstrated that eye closure increased TP across all child groups (all p < 0.0001), but not in adults (p = 0.256). The age × BoS interaction (F(6,425) = 17.25, p < 0.001, η²=0.20) showed that single-leg stance elevated TP at all ages except adulthood. The age × CDT interaction (F(6,425) = 2.19, p = 0.043, η²=0.03) revealed CDT-related reductions in all groups except 7–8-year-olds and adults. A significant gender × CDT interaction further indicated that CDT decreased TP in boys (p < 0.001) but not in girls (p = 0.378). Persistent entropy Persistent entropy showed no significant main effect of vision or any age × vision or gender × vision interactions (all p > 0.05). In contrast, BoS exerted a large effect (F(1,425) = 1084.2, p < 0.001, η²=0.72), with single-leg stance increasing entropy by approximately 8.5%. CDT also produced a smaller but significant increase (F(1,425) = 16.98, p < 0.001, η²=0.05), raising entropy by roughly 1.1%. Age demonstrated a strong effect (F(6,425) = 28.14, p < 0.001, η²=0.28): children aged 7–9 exhibited the highest entropy values, whereas adults showed the lowest (all p < 0.001). Neither the main effect of gender nor the age × gender interaction reached significance. A significant age × BoS interaction (F(6,425) = 8.28, p < 0.001, η²=0.11) indicated that single-leg stance increased entropy consistently across all age groups. The age × CDT interaction (F(6,425) = 3.55, p = 0.002, η²=0.045) revealed CDT-related increases only at ages 8, 10, and 11. Additionally, a significant gender × CDT interaction (F(1,425) = 12.20, p = 0.001, η²=0.03) showed that CDT increased entropy in girls (p < 0.001) but not in boys. H1_NumLongLoops Visual condition did not exert a significant main effect on H1_NumLongLoops (p > 0.05). In contrast, BoS showed a pronounced effect (F(1,425) = 700.83, p < 0.001, η²=0.62), with single-leg stance increasing loop counts by approximately 50%. This effect was further modulated by significant age × BoS (F(6,425) = 11.42, p < 0.001, η²=0.04) and gender × BoS (F(1,425) = 5.89, p = 0.016, η²=0.014) interactions, indicating robust increases across all age groups except adults, and in both sexes except adult males. CDT showed no significant main effect and no gender × CDT interaction; however, the age × CDT interaction was significant (F(6,425) = 2.63, p = 0.016, η²=0.036). Post-hoc tests revealed that CDT reduced loop counts in adults (p = 0.037), increased them in 10-year-olds (p = 0.001), and decreased them in girls (p = 0.036) but not in boys (p = 0.793). Age also produced a significant main effect (F(6,425) = 9.63, p < 0.001, η²=0.12). Adults consistently exhibited the lowest number of long-lived loops, with significant differences relative to all younger groups. Although the age × gender interaction was not significant (F(6,425) = 1.60, p = 0.190, η²=0.02), post-hoc analyses indicated that both adult men and women had fewer loops than all child groups. Additionally, 12-year-old girls showed significantly lower loop counts than other child ages, with values approaching those of adult females. Exploratory correlation analyses between TDA-derived metrics and conventional sway measures are presented in the Supplementary Material (Table S1A–S1H). Across conditions, magnitude-related TDA metrics (e.g., total persistence and Wasserstein distance) showed strong associations with traditional variables such as path length and mean velocity. In contrast, structural descriptors (e.g., number of long loops and persistent entropy) demonstrated weaker and more variable relationships with conventional measures. Discussion The present findings suggest that TDA-derived metrics capture aspects of postural sway that may be relevant for characterizing its geometric structure, although further work is needed to establish their interpretability and physiological significance. Importantly, the present study is primarily designed to examine developmental changes within childhood (7–12 years), with the adult group included as a reference rather than as a continuation of the developmental trajectory. Within this framework, strong age-related differences emerged across the child groups, along with robust effects of visual input, base of support (BoS), and cognitive load. These findings collectively indicate that the topology of sway—its loop structure, persistence characteristics, and entropy—captures features of postural control that are not fully reflected in traditional linear measures [37, 38]. A prominent result was that H1_Wasserstein distance and H1_TotalPersistence both declined markedly with increasing age. Children—especially those aged 7–9 years—showed larger, more H1_NumLongLoops than adults, indicating more oscillatory and less efficient postural corrections. In contrast, adults exhibited persistence diagrams tightly concentrated near the diagonal, reflecting highly damped sway with minimal recurrent cycles. This developmental pattern is consistent with theories of postural maturation suggesting a transition from underdamped, multi-oscillatory control in early childhood toward more critically damped and stable control in adolescence and adulthood [3]. The age profiles of our TDA metrics align well with previous studies reporting greater instability, larger sway amplitudes, and less efficient feedback control in younger children, based on both conventional and nonlinear analyses [2, 4]. Here, TDA provides a geometric characterization of these observations: compared with adults, children exhibited persistence diagrams containing a greater number of long-lived H₁ features, whereas adults displayed fewer and less persistent topological structures. Gender differences emerged but were restricted to the youngest groups. Boys showed slightly higher H1_Wasserstein distance and TP than girls at ages 7–8, but these differences disappeared by later childhood, consistent with reports that girls often develop balance and fine motor control earlier than boys [39]. Importantly, Persistent Entropy did not differ significantly between genders, suggesting that although the amplitude of oscillatory behavior is higher in young boys, the underlying complexity or irregularity of sway patterns is similar across sexes. Thus, gender-related effects in balance appear to reflect magnitude differences rather than fundamental differences in sway dynamics. Contrary to what might be expected from traditional sway measures, removing visual input did not influence Persistent Entropy or the H1_NumLongLoops, despite producing large increases in H1_Wasserstein distance and TP. This contrast indicates that although vision stabilizes posture by reducing sway magnitude, it does not qualitatively alter the structure or complexity of sway. Instead, the geometric structure of the CoP trajectories appears to remain largely preserved, suggesting that visual deprivation primarily affects the amplitude of sway rather than its overall spatial organization. Prior nonlinear studies have similarly reported that eye closure increases sway area but does not consistently change the complexity or regularity of sway patterns in healthy individuals [11, 13]. The stability of entropy and loop count across vision conditions suggests that these TDA metrics reflect inherent sensorimotor coordination properties that are robust to short-term sensory changes. BoS produced some of the largest effects in our study. single-leg stance increased all magnitude-based TDA metrics (H1_Wasserstein distance, TP) substantially, with stronger effects in children than in adults. Persistence diagrams for single-leg stance often showed one or more pronounced loops, indicating dominant oscillatory modes engaged during this challenging task. Young children showed multiple H1_NumLongLoops, consistent with repeated corrective oscillations and difficulty maintaining equilibrium. Adults sometimes displayed a single dominant loop, potentially corresponding to a controlled ankle-strategy oscillation. These findings underscore the sensitivity of TDA metrics to biomechanical difficulty and highlight the potential of loop-based descriptors for quantifying the stability of postural strategies. The cognitive load produced more subtle changes. For H1_Wasserstein distance and TP, dual-tasking reduced or modestly increased values depending on age, with children generally showing more pronounced changes. Persistent Entropy showed a small positive effect of cognitive load, especially in specific child age groups and in girls. This aligns with evidence that attentional demands modulate sway more strongly in children than adults because children rely more heavily on attentional resources to maintain balance [40]. Yet, similar to the vision effect, the dual task did not fundamentally reorganize the topological structure of sway. Rather, it modulated the intensity of existing oscillatory modes. These results support the view that quiet standing in adults is largely automatic and resilient to cognitive interference, whereas children’s balance is still partially attention-dependent. In contrast to magnitude-based metrics, Persistent Entropy exhibited a different developmental trajectory. Entropy declined with age but less steeply than persistence or H1_Wasserstein measures, and under some conditions older children showed entropy values similar to adults. This suggests that while the amplitude and duration of sway oscillations decrease markedly during maturation, the overall distributional complexity of sway remains moderately high even in proficient balancers. This interpretation aligns conceptually with the “optimal movement variability” framework, which proposes that healthy motor systems maintain a certain level of complexity to remain adaptable [41]. Too little variability indicates rigidity, whereas too much indicates noisy or unstable control. The entropy values observed in older children and adults likely reflect this desirable intermediate range. In younger children, however, higher entropy may reflect less structured control rather than optimal adaptability. A notable contribution of this study is the demonstration that TDA captures persistent geometric structures within CoP trajectories that are not directly quantified by conventional sway measures [15]. In the present framework, H₁ features represent loop-like structures that persist across multiple filtration scales and therefore characterize recurring geometric organization within the sway trajectory [14, 16]. Although these features may be influenced by underlying sensorimotor control processes, the current analysis does not permit direct attribution to specific physiological mechanisms. Consequently, the observed topological patterns should be interpreted as descriptors of sway organization rather than direct measures of neuromuscular function or control strategies. For example, a greater number of long-lived H₁ features may reflect more pronounced recurring geometric patterns within the CoP trajectory. In adults, the disappearance or shrinking of such loops is consistent with improved sensorimotor integration and damping. Although we did not perform frequency-domain analyses, previous work indicates that healthy adults exhibit low-frequency (< 0.5 Hz) sway oscillations that might be reflected in the small number of loops detected in their diagrams [22, 42]. Future studies combining TDA with spectral analyses or model-based estimations could clarify the exact physiological mechanisms underlying the detected topological features. Taken together, the present results highlight important differences in how the four TDA metrics contribute to the interpretation of postural control. Importantly, the four TDA metrics did not contribute equally to the interpretation of postural control. A clear distinction emerged between magnitude-based descriptors (H1_Wasserstein distance and H1_TotalPersistence) and structural descriptors (H1_NumLongLoops and Persistent Entropy). The magnitude-based metrics showed the strongest and most consistent effects across age, visual condition, and base of support, and were also strongly correlated with conventional sway measures. This is consistent with the observed correlation patterns, where Wasserstein distance and Total Persistence showed strong associations with traditional sway variables, whereas the other metrics demonstrated weaker or more variable relationships. This pattern indicates that they primarily capture the overall scale of postural variability and are sensitive to task difficulty and developmental stage. In contrast, H1_NumLongLoops and Persistent Entropy provided more selective and context-dependent information. These metrics showed weaker and less consistent associations with traditional sway variables, suggesting that they capture aspects of the geometric organization of CoP trajectories that are not reflected in standard measures. For example, H1_NumLongLoops appeared particularly sensitive to the presence of dominant oscillatory patterns under challenging stability conditions, whereas Persistent Entropy reflected changes in the distributional structure of sway with more subtle sensitivity to task manipulations. Overall, these findings indicate that while Wasserstein distance and Total Persistence are more informative for detecting global changes in sway magnitude, H1_NumLongLoops and Persistent Entropy provide complementary insight into the structural organization of postural control. However, the latter metrics exhibited lower sensitivity to some experimental manipulations, suggesting that their interpretability may be more context-dependent. These findings suggest that not all TDA metrics contribute equally in practical applications, and that metric selection should be guided by the specific research question and sensitivity to the effects of interest. To better contextualize the present findings within established biomechanical frameworks, exploratory correlation analyses were performed between TDA-derived metrics and conventional sway measures. These analyses revealed a consistent pattern across conditions in which magnitude-based TDA features, such as total persistence and Wasserstein distance, were strongly associated with traditional variables like path length and mean velocity, indicating partial redundancy. In contrast, structural descriptors such as the number of long loops and persistent entropy exhibited weaker, inconsistent, and sometimes condition-dependent relationships with classical measures. This suggests that these features capture complementary aspects of postural control related to the geometric organization and coordination structure of sway. Notably, in some task conditions, these relationships were modulated, highlighting the context-dependent nature of TDA-derived descriptors. Overall, these findings support the interpretation that TDA provides both overlapping and novel information beyond conventional biomechanical metrics. These findings have important practical implications. TDA metrics may serve as sensitive markers of balance development and could complement traditional clinical assessments by revealing how, not just how much, individuals sway. Normative topological profiles across childhood could allow clinicians to identify children with delayed or atypical postural control. For research, TDA could support classification of balance impairments, track rehabilitation progress, or detect subtle deficits that linear measures overlook. Given that TDA is robust to noise and captures global geometric structure, its adoption could enhance the precision of computational analyses of balance. This study has several limitations. First, the cross-sectional design limits inferences about individual developmental trajectories; longitudinal studies are needed to better characterize changes in postural control over time. Second, the sample consisted exclusively of healthy participants, and extending this approach to clinical populations would help establish its diagnostic and prognostic relevance. Third, the threshold used to define H1_NumLongLoops, although data-driven, remains somewhat arbitrary; future work could employ more formal approaches such as bootstrapping or persistence-based statistical frameworks to improve robustness. In addition, CoP trajectories were analyzed in their original measurement space without spatial normalization for anthropometric characteristics or base of support. Given that TDA directly captures geometric structure, part of the observed differences may reflect scale-related factors rather than purely functional aspects of postural control. Future studies should therefore explore normalization strategies to better disentangle geometric organization from anthropometric influences. Finally, although TDA provides informative summaries of the geometric and topological structure of CoP trajectories, linking these features to underlying physiological mechanisms will require integration with biomechanical modeling and complementary analytical approaches. Conclusion In conclusion, this study demonstrates that Topological Data Analysis offers a sensitive and conceptually meaningful framework for quantifying developmental changes in postural control. By examining the geometric structure of CoP trajectories, we showed that younger children exhibit a greater number of H1_NumLongLoops —reflecting oscillatory and less efficient control—while older children and adults display topologically simpler and more stable sway patterns. Metrics such as H1_Wasserstein distance and TP proved especially effective in capturing these developmental differences and in reflecting task-related instability. Conversely, Persistent Entropy and H1_NumLongLoops counts were relatively stable across visual and cognitive manipulations, suggesting that they index intrinsic features of an individual’s coordination strategy rather than short-term sensory effects. Together, these findings highlight the potential of TDA-derived measures as objective biomarkers for tracking balance maturation, identifying atypical developmental patterns, and monitoring rehabilitation progress. By integrating mathematical topology with biomechanics, this work underscores TDA as a valuable addition to current movement-science methodologies. Future research should extend this approach to clinical and older adult populations, as well as dynamic balance tasks, to further establish the relevance of topological descriptors for understanding neural and mechanical mechanisms of postural control. Supplementary Information Supplementary Material 1. Acknowledgements We thank the participating students and their families for their cooperation throughout the study. We also acknowledge the assistance of the research staff at the Biomechanics Laboratory of Islamic Azad University, Hamedan Branch, for their support during data acquisition and logistical coordination. Declaration of generative AI and AI-assisted technologies in the writing process During the preparation of this work, the authors used ChatGPT, an AI-assisted tool, to improve the readability and language of the manuscript. After using this tool, the authors carefully reviewed and edited the content as needed and take full responsibility for the content of the published article. Authors’ contributions E.A. contributed to study design, participant recruitment, data collection, and preliminary data processing. M.M. (corresponding author) was responsible for developing the analytical pipeline, performing all statistical analyses, interpreting the findings, and drafting and revising the manuscript. Both authors reviewed and approved the final version of the manuscript. Funding This research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors. Data availability The datasets generated and analyzed during the current study are available from the corresponding author upon reasonable request. Declarations Ethics approval and consent to participate The study was approved by the Ethics Committee of Islamic Azad University, Hamedan Branch (Approval Code: IR.IAU.H.REC.1400.002). Written informed consent was obtained from all participants or their legal guardians prior to data collection. Consent for publication Both authors have reviewed the manuscript and consent to its publication. 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. Data Availability Statement The datasets generated and analyzed during the current study are available from the corresponding author upon reasonable request.