Sex and age differences of postural control in community-dwelling older adults

Postural stability was assessed in 4,588 elderly participants from the AGES-Reykjavik Study cohort, aged 65–95 years in connection with a larger study (Age gene/Environ-ment Susceptibility Reykjavik Study; AGES-Reykjavik). The cohort was randomly sampled from 30,795 Reykjavik residents born between 1907 and 1935. Recruitment achieved a rate of 62%, with 19,381 individuals attending the initial assessment. By 2006, 5,764 surviving members were re-examined, and 4,588 participants (2,664 females and 1,925 males; mean age 76.2 ± 5.4 years) were included in this study. Figure 1 shows the age distribution of the subjects. Participants were stratified into five age groups: 65–69, 70–74, 75–79, 80–84, and ≥85 years.

Exclusion criteria included blindness, wheelchair dependency, inability to stand unaided for 30 s, or requiring assistance during a chair rise test. Ethical approval was obtained from the Icelandic National Bioethics Committee (VSN: 00–063) (Figure 2).

Postural stability was measured using a custom-built static force platform (Aalto et al., 1988b) that recorded the center point of force (CPF) location that calculates the changes of force.

(or pressure) on platform surface. Participants stood with arms crossed over chest (if possible), knees locked and maintained stability (Figure 3). Two conditions were tested: eyes open for 30 s and eyes closed for 30 s. Only eyes-closed conditions are reported apart of the Rombergs quotient. For referee measurements we used Romberg’s quotient to illustrate robustness of this parameter in age and sex evaluation.

Stabilogram data was sampled at 50 Hz, captured with medio-lateral and anterior–posterior force components and total reaction force. To remove transient artifacts the signal was filtered with median filter (Aalto et al., 1988a). Signal contamination from electrical and biological noise was mitigated using a Chebyshev finite impulse response (FIR) low-pass filter (17 Hz passband, 21 Hz stopband, maximum passband ripple 1 dB, stopband attenuation 80 dB). When subjects were placed on the force platform, they often tend to readjust their position after beginning of the measurement leading to transition error (Peterka, 2000). Preprocessing excluded the initial 15 s of data at the beginning of signal, leaving 600 samples (12 s) for analysis. Before further processing, we removed the mean values of the stabilogram’s positional components. The weight signals were normalized, altering the mean to 0 and the variance to 1. However, in the moment calculations we used the filtered weight signal without normalizing.

The selection of the variables for the analysis was carried out following the procedure reported here below.

In the final analysis we focused on following six time-domain variables (Pyykkö et al., 2000; Toppila and Pyykkö, 2000; Rasku et al., 2012) (Table 1).

Variable C(Y). Using positional data, we calculated the peak-to-peak sway amplitudes in the anterior–posterior direction and determined the confidence limits encompassing 95% of the stabilogram samples. These metrics represent the 95% confidence amplitude projections in the anterior–posterior direction, denoted as C(Y).

Variable M(MY). Corrective movements influence the stabilogram by inducing changes in moment and force reactions due to ankle and hip torque adjustments. To quantify these effects, we calculated the mean absolute moment about the anterior–posterior axis, normalized by the subject’s body mass. This metric is referred to as M(MY).

Variables ZCR(Y) and ZCR(VX). To characterize postural force activity and quantify postural instability, we measured the frequency of CPF crossings over the stability center point. In the medio-lateral direction we calculated how many times the center point of force crosses the y-axis ZCR(X) and in antero-posterior direction ZCR(Y). We also calculated how many times the velocity changes direction in the mediolateral direction ZCR(VX). This frequency is defined as the zero-crossing rate either by sway in antero-posterior direction, denoted as ZCR(Y) or by velocity in the medio-lateral direction, denoted as ZCR(VX). Large values of zero crossings indicate that a person’s swaying amplitudes are small, and they occur around a point that remains sufficiently stationary. On contrary, in small zero crossing situations there are long swaying amplitudes or many points around which the swaying occurs.

Variable CRI(T). To assess the efficiency of postural control, we determined the critical time [CRI(T)] marking the transition from an open-loop to closed-loop control dynamics, following the method outlined by Collins and De Luca (1993, 1995). This metric has been widely used to infer the coexistence of two mechanisms during quiet stance: an open-loop mechanism operating over short time intervals, and a closed-loop mechanism dominating at longer intervals. CRI(T) determines the time when open- loop mechanism is changing to closed-loop mechanism.

Additionally, we identified steady stance periods by calculating the number of samples corresponding to steady-phase standing intervals [ST(N)]. Figure 4 illustrates examples of presence of steady phases during quiet standing. Steady-phase standing periods [ST(N)] were identified by summing the moving variances of the moments about the medio-lateral and anterior–posterior axes. This summed moving-variance signal was normalized by dividing it by the square of the subject’s mass to mitigate mass-related influences. The threshold of low variation was set to 0.002 corresponding to the upper limit of the lowest quartile of mean value of variance of weight signal over all subjects. This value was compared to the value of the moving variance. If the moving variance was less than or equal to 0.002, the respective sample in the centered weight signal was designated as belonging to a “low variation” period.

A moment signal sample (Mi) was classified as part of a steady phase if the local variance of five consecutive moment samples, normalized by the square of the subject’s mass (Rasku et al., 2012). The local variance about a single axis at a given point (Mi) was calculated as described in formula 1. The moment reactions occur when a correction movement is made.

The 5-point moving variance centered at pointis: i

Where the local mean is.

M¯ is the moving average of 5 moment samples, and the variance is calculated with the two preceding and two following samples of M_i.

To elucidate the classical analysis of posturography the Romberg quotient (RQ) was calculated for AP-sway velocity and for peak-to peak AP sway path.

For the Romberg’s quotient (RQ) analysis, we used SPSS version 26. Linear regression was applied to examine whether age or sex could be predicted from the RQ values. Finally, a decision tree model was implemented. In constructing the model, each group of 100 participants served as a training subset, and the final group of participants was used for model validation.

The data was stratified to approximate normal distributions. Natural logarithmic transformations were applied to the parameters. In data processing and statistical analysis MATLAB ver. R2014a was used. To predict sex with decision tree algorithm, the dataset was balanced by selecting 1,925 females whose age distribution matched that of the male participants. The male and female data were randomized and combined into a single data matrix, arranged in an alternating sequence of male and female samples. This matrix was then divided into 10 mutually exclusive subsets.

In modelling a tenfold cross-validation method was employed for decision tree analysis. In each iteration, nine subsets were used to train the model and estimate its parameters, while the remaining subset was used as the test set to evaluate the model’s predictive accuracy. During each round of classification, predictions were generated for 192 males and 192 females. The accuracy of sex classification was evaluated via repeated cross-validation. Cross-validation is a model validation technique used in machine learning to assess how well a model generalizes unseen data.

The RQ was measured in peak-to-peak antero-posterior (ppAP) sway path and in antero-posterior sway velocity. The correlation between velocity and sway path was 0.618 in visual and nonvisual conditions indicating that these variables contain to great extent the same information. The measured values of visual and nonvisual condition in males and females are shown in Appendix. The females swayed less both in visual and non-visual condition than male. The sex difference in Romberg’s quotient was prominent (Mann–Whitney test p < 0.001 for both ppAP sway path and RQ-sway velocity). The females were less influenced by vision than males (Figure 5). There were significant differences in male between the RQ of ppAP sway path and RQ-sway velocity (Wilcoxon test, p < 0.001) but not in female. The RQ-sway path and sway velocity could predict the sex correctly with accuracy of 60.4 per cent (logistic regression analysis, p < 0.001). In decision tree analysis the sex could be predicted correctly in 14 percent of females and 92 percent of males providing overall percentage to 60 percent.

We observed significant differences between age groups in the RQ of ppAP sway path analysis (ANOVA, F = 5.73, p < 0.001), but not in the RQ of antero-posterior sway velocity (ANOVA, F = 1.40, p = 0.231). Pairwise comparisons revealed differences in RQ ppAP sway path and anteroposterior sway velocity analyses (Figure 6). These results highlight the statistical robustness of the models, with strong significance and moderate explanatory power.

However, substantial overlap between age groups was observed, which limits the ability to classify individuals-based age groups. In decision tree analysis the age group discrimination with accuracy of 5 years division provided an overall accuracy of 30.5 percent in both variables.

Table 2 shows the mean outcomes of posturography measurements for males and females. The final column (N) indicates the number of participants in each age group. These variables were selected based on best prediction of differentiation between sex and age groups. As an example, Figure 7 illustrates the linear trend of M[MY] with increasing age. The mean values are plotted separately for males and females, with standard deviations included for each sex. The plot reveals that males rarely scored below 0.2, while females seldom exceeded 0.5. This suggests that muscular strength plays a critical role in postural correction, as males generally possess greater muscular force, enabling more effective compensation for poor postural control.

Table 3 presents the correlation matrix of the measured parameters. The matrix was computed using a dataset that includes both male and female participants. Although the correlation coefficients showed minor sex-specific differences, the overall structure of the matrix remained consistent across groups. The abbreviations for all variables are provided in Table 1.

Table 4 summarizes the statistical differences in individual variables between males and females. In the indicator column, a value of 0 denotes no statistically significant difference, whereas a value of 1 indicates that the variable differs between groups (Wilcoxon test, p < 0.05).

The amplitude range of anterior–posterior sway [C(Y)] was identified as the most effective parameter for predicting age. This parameter demonstrated statistically significant differences across all successive age groups for both males and females. Tables 5, 6 present the differences of variables between successive age groups among males and females.

Although not all parameters exhibited statistically significant differences between consecutive age groups, they followed consistent trends. The mean moment in anterior–posterior sway [M(MY)] increased with age, reflecting greater sway and reduced postural stability.

Clinically, M(MY) may serve as a valuable screening tool for evaluating postural compensation, particularly in cases of vestibular lesions or among frail older individuals (Tuunainen et al., 2011). This suggests that muscular strength plays a critical role in postural correction.

Age prediction was conducted using data vectors comprising C(Y), ZCR(Y), CRI(T), and ST(N). For males, the standard deviation of the age prediction error was 4.9 years using linear models and 5.0 years with classification trees. Similarly, for females, the standard deviations were 5.0 years and 5.2 years, respectively.

In the linear models for predicting age for males were R² = 0.152, F = 62.0 and p < 0,01, respectively. For females the respective values were R² = 0.137, F = 75.9, and p < 0,01, respectively. These results highlight the statistical robustness of the models, with strong significance and moderate explanatory power.

Figure 8 depicts the decision tree, where leaf nodes represent sex. The input data vectors for sex prediction included C(Y), M(MY), ZCR(VX), CRI(T), and ST(N). The classification tree achieved an overall sex recognition accuracy of 71 ± 2%, with 28 ± 4% of males misclassified as females and 29 ± 4% of females misclassified as males.

Notably, data for ZCR(VX), CRI(T), and ST(N) were excluded from the final model, indicating that these parameters were redundant for sex prediction. This suggests that sex classification is primarily influenced by other metrics such as C(Y) and M(MY), highlighting potential avenues for refining predictive models.

Collins and De Luca (1993) introduced stabilogram diffusion analysis, a method that provides a quantitative statistical characterization of the apparently random variations in CPF trajectories recorded during quiet upright stance in humans. This approach produces a stabilogram diffusion function (SDF), which describes the mean square CPF displacement as a function of the time interval between CPF samples.

SDFs typically exhibit a biphasic form, suggesting the involvement of two distinct control regimes: a short-term open-loop process and a longer-term closed-loop Collins and De Luca (1993) also defined the critical point, corresponding to the time interval and SDF value at which the short-term and long-term linear fits intersect. This metric has been widely used to infer the coexistence of two control mechanisms during quiet stance: an open-loop mechanism operating over short time intervals, and a closed-loop mechanism dominating at longer intervals.

Subsequent studies have established normative values for SDF parameters (Collins and De Luca, 1993) and have demonstrated systematic changes in these parameters as a function of visual input (Collins and De Luca, 1995; Riley et al., 1997) age (Collins et al., 1995), and pathological conditions (Treger et al., 2020). In the present study, SDF parameters were applied to estimate the critical time of transition between open- and closed-loop postural control [CRI(T)].

A decline in the critical time for transitioning between open- and closed-loop postural control [CRI(T)] was observed with advancing age. Open-loop control relies on pre-planned strategies independent of environmental feedback, whereas closed-loop control incorporates sensory information from the vestibular, visual, and somatosensory systems. A reduced CRI(T) reflects a diminished ability to predict prior positions, resulting in prolonged sway and an increased risk of falls. These findings are consistent with previous reports describing the loss of “recent postural position memory” (Pyykkö et al., 2000). Fear of falling may further exacerbate these effects by restricting natural sway and limiting movement.

CRI(T) analysis also indicated unstable short-term behavior, interpreted as greater sway distance before the initiation of closed-loop control. A significant group-by-condition interaction revealed that older adults exhibit a greater reliance on visual input compared with younger, healthy individuals.

Later work demonstrated that similar results can be obtained using a linear fully closed-loop system (Peterka, 2000). In this model, the body is represented as an inverted pendulum with torque applied at the ankle joints. The applied torque consists of both a random disturbance component and a control component. The control torque is subject to a time delay that reflects neural conduction, processing, and muscle activation delays. This modeling framework enables interpretation of experimentally observed stabilogram changes in terms of variations in neural controller parameters and time delays, rather than simply distinguishing between open-loop and closed-loop control. This aspect of parameter selection remains to be studied.

Two additional parameters, the zero-crossing rate of mediolateral sway velocity [ZCR(VX)] and the steady-phase standing samples [ST(N)], were strongly influenced by age. Individuals younger than 75 years demonstrated a sharp increase in ZCR(VX) accompanied by a marked decrease in ST(N), which may reflect reduced physical activity levels (Melzer et al., 2003). Beyond the age of 75, changes in these parameters became more gradual but continued to follow a declining trajectory.

Analysis of ZCR(VX) and ZCR(Y) is particularly informative, as transient features in stabilogram signals are strongly associated with postural control strategies and compensatory actions (Prieto et al., 1996). Rehabilitation programs have been shown to “refresh postural memory,” thereby reducing transients and enhancing the effective use of the stability area (Tuunainen et al., 2013). These findings highlight the importance of ST(N) and ZCR(VX) in posturography for evaluating postural control and guiding the development of targeted rehabilitation strategies.

Aging influenced postural control, with the amplitude range in the anterior–posterior direction [C(Y)] identified as the most robust predictor of age as has been indicated previously (Pyykko et al., 1990; Fujita et al., 2005; Johansson et al., 2019). The age was predicted with an accuracy of ±5 years. We refer to this amplitude control, in that long swaying amplitudes were pronounced, allowing the person to react when postural confidence limits are reached. Normalizing the logarithm of the 95% anterior–posterior sway amplitude by body mass identified that this parameter explained approximately 15% of age-related variability. Studies have linked increased body sway with higher fall risk, especially among older women (Pajala et al., 2008; Rezaei et al., 2024).

From a biomechanical perspective, maintaining posture requires the CPF to stay within the base of support. Sensory deficits in aging, such as diminished sensitivity of foot mechanoreceptors and proprioception in calf muscles, shift reliance to visual and vestibular systems (Toppila and Pyykkö, 2000; Hytönen et al., 1993). This reliance likely explains the increased sway amplitudes seen with age and the corresponding elevated fall risk (Pyykkö et al., 2000). Strategies that maximize sway range may provide additional sensory feedback but can lead to static muscle loading and fatigue. These shifts in the CPF neutral position manifest as random positional changes, reflected in parameters such as ZCR(Y), ZCR(VX), and ST(N). In this postural control, a high oscillation rate about the stationary point was the dominant characteristic of maintaining postural control. The zero crossing rates [ZCR(Y) and ZCR(VX)] and critical time CRI(T) during which an open loop control changes into a closed loop provide information about how the subject is gaining balance control. A short critical time period CRI(T) very rapidly controls excessive postural oscillations and appears to be effective strategy in balancing a human.

Consistent with prior studies, males exhibited greater sway amplitudes and moments, while females demonstrated superior stability (Julienne et al., 2024; Era et al., 2002). The mean moment in antero-posterior sway [M(MY)] and steady-phase standing periods [ST(N)] were the most predictive sex-specific parameters. These results suggest that males employ higher-moment strategies requiring greater energy and muscle activity, whereas females rely on more stable postural control mechanisms.

In the pendulum model used to interpret balance control in posturography, several biomechanical factors influence postural performance. These can be explored by identifying variables that capture different components of the stabilogram signal, which may have clinical relevance. Important targets include fall prevention in the elderly and the characterization of postural impairment in vestibular patients. However, no consensus currently exists regarding the optimal set of parameters (Kingma et al., 2011). This lack of agreement has hindered the widespread clinical adoption of posturography.

In the present study, we selected a comprehensive set of variables that quantify different aspects of postural stability and by evaluating their predictive value on age and sex by using artificial intelligence (AI). Pennone et al. (2024) attempted to predict falls in older adults using 53 static posturography parameters combined with AI. Their prediction accuracy was limited to 60%, and the study did not report the specific decision tree branches or leaves contributing to this outcome. Cabral et al. (2020) applied multiple AI techniques to a large dataset of community-dwelling older adults (aged 60–88 years) in a cohort expected to improve predictive sensitivity. They found that static posturography did not enhance the prediction of single or recurrent falls. Sarmah et al. (2024) used neural networks with sway-path parameters and obtained similar results, although they noted that balance analysis remains valuable for the early detection of balance deficits, facilitating timely medical care and improving quality of life through rehabilitation.

We put much effort into choosing parameters describing information on postural control strategy. We assume that variables exhibiting strong internal correlations (e.g., sway path and sway velocity) reflect the same underlying phenomenon and therefore jointly provide limited additional information on postural control strategies, in contrast to metrics that are uncorrelated or only weakly correlated. However, this assumption may be questioned, as both sway velocity and sway path might potentially capture distinct aspects of postural control.

In their Cochrane review of posturography outcomes, Alonso et al. (2012) highlighted a scarcity of anthropometric data across the 17 studies included for sex- and age-specific comparisons, comprising 5,194 participants. Only a few studies reported key variables such as height and weight, even within so-called “normative” datasets (Lelard and Ahmaidi, 2015). Similar limitations were noted in studies evaluating the effects of proprioceptive or dynamic training on postural stability in older adults (Chiari et al., 2002). In several original reports, age and sex were not analyzed, as these factors were considered non-essential to the primary outcomes.

Among anthropometric variables, height exerted the greatest influence on postural balance, both in the overall sample and when examined separately by sex. Moreover, postural balance was more strongly influenced by anthropometric factors in males than in females (Alonso et al., 2012). A crude estimate suggests that every additional 10 cm of height increases sway amplitude and velocity by approximately 5–10%, and very tall individuals may sway 15–20% faster than shorter individuals, even when their balance is otherwise normal. In contrast, weight accounts for only 1–3% of the variability in sway velocity. In the study by Wiegmann et al. (2020), the mean height difference between men and women was only 5.8 cm, suggesting that height may be a relatively minor confounder.

In the present study, height data were not available at the time of data collection and therefore could not be incorporated into standardization procedures. Height may influence sway amplitudes, velocities, and moment-related measures, but it is unlikely to affect stabilization time or estimates of critical time. Moreover, the literature remains somewhat inconsistent regarding the magnitude of age-related effects on postural stability (Alonso et al., 2012; Wiegmann et al., 2020). Based on meta-analytic evidence, Alonso et al. (2012) reported significantly greater sway in individuals aged 50–79 years compared with younger adults. Including height data might have enabled somewhat finer age-related resolution in the current analysis. However, uncertainties introduced by height are minimized when evaluating the Romberg quotient, as it is expressed as a ratio rather than an absolute measure.

We acknowledge that the exclusion criteria tend to exclude the frailest older adults, who are at the highest risk of falls and for whom balance assessment is interesting. However, apart from individuals with visual impairment, such persons are seldom exposed to balance tests due to recognized elevated fall risk by clinical condition.

In this report, we present data on eyes-closed quiet standing on a force platform, except for the Romberg quotient analysis, as our aim was to focus on the effects of ageing and sex on multiple measures of postural stability. We intend to report results from other balance-demanding tests—such as target-reaching, gait assessment, and chair-rising—in a separate work.

The original contributions presented in the study are included in the article/, further inquiries can be directed to the corresponding author. Supplementary material