Abstract

Objective:

Force platforms are widely used to assess postural stability and fall risk in older adults. However, traditional parameters often capture overlapping phenomena and fail to fully reflect underlying control mechanisms. This study evaluated combining of six partly independent parameters to distinguish sex and age-related differences in postural control among community-dwelling elderly.

Methods:

A total of 4,588 adults aged 65–95 years were assessed using static posturography under non-visual conditions. Six time-domain parameters, reflecting torque control, positional control and anticipatory control of the center point of force. Romberg’s quotient was included for comparison.

Results:

Females exhibited greater stability, whereas males relied more on corrective force moments and showed larger sway amplitudes. Classification trees predicted sex with 71% accuracy using three parameters. Aging was associated with increased anteroposterior sway amplitude and a reduction in the critical time for transition between open- and closed-loop control. Additional age-sensitive parameters included mediolateral velocity zero-crossing rate and steady-phase duration. Age could be predicted within ±5 years for both sexes. Romberg’s quotient could discriminate age in 30% and sex differences in 60% of participants, only.

Conclusion:

Postural stability is influenced by both sex and age. The identified combination of parameters provides a framework for estimating the “biological age” of postural control and investigate balance impairments. Age-related decline appears consistent within a 5-year range, bur does not exclude the effect of lifestyle or comorbid factors. This study demonstrates that use of multidimensional data vectors with the implementation of AI-modeling can improve the predictive accuracy and clinical applicability of posturography.

Introduction

Adequate muscle strength, efficient gait, and good balance are essential components of independence and overall wellbeing, yet these capacities decline with age (Era et al., 2006). Investigating how aging impacts parameters related to balance and strength is of critical importance for both clinicians and patients. Balance is maintained by sensory inputs from vision, the somatosensory system, and the vestibular system, which work together to detect and process positional, velocity, and acceleration information, thereby maintaining postural stability (Horak, 2006; Pyykkö et al., 1991; Enbom et al., 1991).

Human postural control is dynamic, evolving through context-dependent learning and subject to age-related changes (Abrahamova and Hlavacka, 2008). Declines in postural control typically begin around the age of 60 (Era et al., 2006), increasing susceptibility to stumbles and falls (Tinetti et al., 1994; Maki et al., 1994). Approximately 50% of individuals aged 75 years or older experience fall annually (Fuller, 2000). Fear of falling further restricts daily activity, diminishing quality of life (Tuunainen et al., 2011). The multifactorial etiology of falls (Blake et al., 1988; Cass et al., 1996; Tuunainen et al., 2013; Jäntti et al., 1995) complicates the identification of primary balance deficiencies. Postural stability is commonly evaluated using force platforms, which measure center-of-pressure movement and generate stabilograms (Nashner, 1983). Despite the widespread use of posturography for rapid screening (Era et al., 2006; Maki et al., 1994; Julienne et al., 2024), its clinical utility remains debated (Kingma et al., 2011; Pyykkö et al., 2000; Visser et al., 2008). However, individual sway characteristics depend on context (Horak, 2006; Toppila and Pyykkö, 2000), age (Hytönen et al., 1993), prior experience (Gautier et al., 2008), and overall health (Sack et al., 1993). The complexity of stabilogram-derived data can obscure meaningful insights into balance deficits (Palmieri et al., 2002; Piirtola and Era, 2006).

Haeggstrom et al. (2006) identified 163 distinct parameters used to describe human posture, underscoring the lack of consensus in measurement and analysis (Kingma et al., 2011). Previously we studied interrelations among 30 time-domain variables and showed that velocity- and amplitude-related variables are strongly associated, as are moment- and sway-area-related parameters, reflecting overlapping control mechanisms (Rasku et al., 2012).

The aim of the present study was to determine whether a selected set of stabilogram variables, recorded during quiet stance without visual input, can identify sex- and age-related differences in postural control among community-dwelling older adults. We examined the hypothesis that postural control strategies in the elderly comprise three components: (a) torque control, (b) positional control, and (c) anticipatory control that can be recorded on the force platform.

Materials and methodsParticipants and study design

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.

Participants in the survey and their number.

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).

A subject standing on force platform.

Test setup

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.

Age distribution of the subjects.

Data processing

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.

Selected variables

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

We first analyzed 30 different variables which suggested three different controlling strategies consisting of (1) torque control, (2) positional control and (3) anticipatory control of the center point of force (Supplementary Tables 1–3).

Then we analyzed correlation between the variables and found that the variables could be diminished to 17 without significant information loss (Rasku et al., 2012).

For these remaining variables we performed factor analysis and excluded variables that were associated with only modest loading on all factors. The threshold for the exclusion was 0.1. We also excluded variables with correlations over 0.5 as these were at least partly describing the same control strategy. In addition to the commonly used variables, we included variables that we calculated and considered important. The results of the analysis suggest that three different strategies (factors) are used to maintain an upright stance. We characterized the factors as follows: factor 1 is a torque control, factor 2 a positional control and factor 3 an anticipatory control. These 3 factors accounted for 63% of the variance in the selected variable set (Rasku et al., 2012).

From the final variables we further eliminated less important parameters (see Appendix) that were not associated with unique postural control strategy leaving six time-domain parameters.

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 abbreviationVariable nameVariable description/meaningC(Y)Amplitude range of anteroposterior sway (peak-to-peak)C(Y) represents the range of body sway in the anteroposterior (Y) direction. A larger range indicates greater postural instability.M(MY)Mean moment in anteroposterior swayM(MY) reflects neural conduction, processing, and muscle activation within a pendulum model of posture. Higher values are associated with increased muscle force activity required to maintain postural stability.ZCR(Y)Zero-crossing rate of anteroposterior swayZCR(Y) characterizes postural control responses associated with increased physical effort to regulate body sway across the neutral center point in the anteroposterior direction.ZCR(VX)Zero-crossing rate of mediolateral sway velocityZCR(VX) describes changes in CPF sway velocity across the neutral center point in the mediolateral direction. Higher values reflect postural instability and increased physical effort required to control body sway.CRI(T)Critical time for open- to closed-loop transitionOpen-loop control relies on pre-programmed strategies independent of sensory feedback, whereas closed-loop control incorporates vestibular, visual, and somatosensory information. CRI(T) denotes the critical time point at which short-term and long-term linear fits intersect, corresponding to the transition between these control mechanisms.ST(N)Steady-phase standing periodsST(N) represents local stationary periods during stance and reflects reduced postural muscle activity between transient changes in center-of-pressure behavior.

Variables selected to describe postural control in elderly.

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.

The steady-state periods (round markers) during stance on force plate posturography.

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 point is:

Where the local mean is.

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

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.

Statistical testing and validation

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.

ResultsIllustrating classical measures as Romberg’s quotient between sex and age groups

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.

Romberg’s quotient in male and female when calculated on for antero-posterior mean velocity (dotted line) and peak-to-peak antero-posterior sway (solid line). Mean and 95% confidence intervals are shown.

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.

Romberg’s quotient for different age groups is presented for anteroposterior mean velocity (dotted line) and peak-to-peak antero-posterior sway (solid line). Data are expressed as mean values with 95% confidence intervals.

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.

Evaluation of sex differences with six parameters

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.

AgeC[Y] cmM[MY] kg m/sZCR[Y]
number in measuring periodZCR[VX]
number in measuring periodCRI[T]
s in measuring periodST[N]
number in measuring periodNMales≤690.230.4314.8871.0314117670–740.300.4717.71400.978158875–790.350.4818.71310.918359080–840.400.4919.71240.9282431≥850.490.5322.11230.8366140Females≤690.230.3215.71481.1022531870–740.260.3217.31651.1018380975–790.290.3317.71581.0018276080–840.360.3619.41450.93162594≥850.410.3622.11670.96146182

Mean values of the parameters for both males and females.

C(Y), Amplitude range of the antero-posterior sway; M(MY), Mean moment in antero-posterior sway; ZCR(Y), zero crossing rate of the antero-posterior movement; ZCR(VX), zero-crossing rate of medio-lateral sway velocity; CRI(T), critical time for open and closed loop control; ST(N), Steady-phase standing periods; N, number of subjects in the respective age group.

Mean values and standard deviations of the absolute moments [M(MY)] for males and females.

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.

ParametersC(Y)M(MY)ZCR(Y)ZCR(VX)CRI(T)ST(N)AGESEXC(Y)1.000.460.480.05−0.30−0.580.32−0.13M(MY)0.461.00−0.01−0.35−0.04−0.510.10−0.33ZCR(Y)0.48−0.011.000.33−0.39−0.370.18−0.03ZCR(VX)0.05−0.350.331.000.16−0.11−0.010.19CRI(T)−0.30−0.04−0.390.161.000.27−0.110.08ST(N)−0.58−0.51−0.37−0.110.271.00−0.130.35AGE0.320.100.18−0.01−0.11−0.131.00−0.03SEX−0.13−0.33−0.030.190.080.35−0.031.00

Correlation of the selected variables.

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).

AgeGrC(Y)M(MY)ZCR(Y)ZCR(VX)CRI(T)ST(N)10101012111111311111141101015110111

Statistical differences of variables between males and females in different age groups (AgeGr).

0 indicates no difference and 1 difference at level of p < 0.05. For abbreviations of variables see Table 1.