Postural stability during a longitudinal expedition in an isolated and confined Antarctic environment

Abstract

Postural stability is critical for safety in extreme environments like Antarctica, where falls pose a significant risk to expedition crews. However, systematic longitudinal studies from this setting are lacking. In the first such study, we repeatedly assessed postural control in thirteen members of a 49-day Antarctic mission using a low-cost, autonomous system under four sensory conditions (firm surface vs. compliant surface eyes open vs. eyes closed). We hypothesized that postural stability would initially deteriorate due to acute stress and environmental novelty, followed by gradual recovery and adaptation over the mid-term . Our comprehensive analysis revealed a marked deterioration of stability when visual or proprioceptive input was compromised, manifesting as increased sway and a shift toward less complex, more persistent control dynamics. Contrary to our hypothesis, no systematic longitudinal adaptation was detected. These findings highlight the increased reliance on visual information when proprioceptive cues are unreliable and suggest that postural strategies remain remarkably stable during a medium-term polar mission. Despite its known technical limitations, our work demonstrates the practical feasibility of using this autonomous system for field stabilometry in isolated environments.

Introduction

Antarctica is an ideal setting for studying the effects of isolated and confined environments on the human organism. While distinct from the microgravity environment of spaceflight, the J.G. Mendel Station has been characterized as a terrestrial analog for the Isolated, Confined, and Extreme (ICE) stressors¹. Given that postural control requires significant attentional resources², we hypothesized that the combined effects of cognitive fatigue³ and environmental constraints, such as cold exposure⁴ and altered locomotor strategies on icy terrain⁵, would manifest as an initial impairment in stability followed by gradual adaptation during mid-term (in contrast to the secondary decline often observed in long-duration overwintering missions⁶). A long-term, 14-month expedition in Antarctica has shown that isolation and a monotonous environment can negatively affect brain plasticity³. Typical symptoms associated with polar expeditions include sleep disturbances, cognitive decline, negative moods, and interpersonal tension, with severe psychological disorders occurring in approximately 5% of crew members⁶.

Beyond these well-documented psychosocial stressors, the extreme physical environment imposes unique physiological and biomechanical demands on expedition members. The average temperature remains below freezing even during the Antarctic summer⁷, and the surface consists predominantly of ice and snow, or, in deglaciated areas like the Ulu Peninsula on James Ross Island, a polar desert with a substrate of permafrost overlaid by a 0.5–1.2 m thick active layer^8,9. These variable conditions affect gait parameters. Low-friction surfaces significantly alter step length and walking speed⁵, and adaptive changes in muscle activity occur on icy surfaces. Specifically, during heel contact, activity is reduced in the m. tibialis anterior and m. rectus femoris, while in the mid-stance phase, activity of the m. gastrocnemius lateralis decreases¹⁰. This reflects a more cautious stepping strategy to avoid slipping, characterized by placing the foot more flatly on the surface. Furthermore, exposure to low temperatures (10 C) alone can significantly increase postural sway, specifically parameters such as sway path length (by 67–87%) and total sway area (by 42–67 %)⁴.

Despite the clear relevance of postural control for safety and performance in this environment, a notable gap exists in the scientific literature. To our knowledge, no studies have published results from stabilometric measurements conducted directly in Antarctica. There are only indirect mentions, for example, of the inclusion of posturography in biomedical experiments at Concordia Station¹¹ or a description of testing with a force plate as part of the ESA BoneHealth project¹². However, these sources do not provide published results. In Antarctica, postural stability is of immediate practical importance, as fall prevention¹³ is critical in an environment with limited access to healthcare¹⁴.

To address this research gap, we employed the Wii Balance Board (WBB, Nintendo Co., Ltd., Japan) for the systematic assessment of postural stability. The validity and reliability of this tool for measuring standard Center of Pressure (CoP) parameters have been well-established^15,16. The suitability of the WBB for our study is further supported by its potential for advanced analyses. As recently shown by Hua et al.¹⁷, data from the WBB allow for a frequency-based decomposition of the signal into Rambling (RM) and Trembling (TR) components¹⁸, which expands its diagnostic capabilities for field research.

Our study had three key aims: (i) to explore the hypothesis that postural stability changes over the course of an Antarctic expedition. Specifically, we formulated an exploratory, biphasic hypothesis that stability would initially deteriorate upon arrival due to acute stress and unfamiliarity with the surface, followed by gradual motor adaptation reflecting physiological acclimatization and learning. While these directional hypotheses primarily concerned linear measures of sway magnitude, the analysis of nonlinear parameters (e.g., Fractal Dimension, Lyapunov Exponent) was conducted as an exploratory aim. We sought to investigate whether the temporal structure of the sway signal, which reflects control complexity and adaptability, might reveal subtle sensorimotor changes that are not captured by traditional time-domain metrics. (ii) to assess the diagnostic value of different parameter classes (time-, frequency-, and nonlinear-domain, including RM/TR decomposition and Romberg quotients) in distinguishing the effects of visual feedback and surface type, and (iii) to evaluate the field feasibility and technical performance of the WBB-based system under Antarctic conditions. This study presents the first longitudinally structured stabilometric dataset from Antarctica. It provides preliminary mechanistic insights that may inform diagnostic screening for fall risk in extreme environments and contributes to the understanding of human adaptation in isolated and extreme environments.

Methods

Mission

The expedition was conducted during the 2025 austral summer at the J. G. Mendel Czech Antarctic Station (JGM), located on the northern coast of James Ross Island, near the Antarctic Peninsula. For 49 days, a 14-member international and interdisciplinary team was based at the station. The primary objective of the mission was comprehensive multidisciplinary research focusing on three main areas: (i) human physiological and psychological adaptation to an isolated and extreme environment, (ii) environmental sciences, and (iii) new technology validation.

The site is characterized by a relatively arid and windy polar maritime climate. The deglaciated landscape (Fig. 1) experiences mean annual air temperatures of approximately –5.8 C, though summer averages often rise above freezing⁷. The station is designed to accommodate 15–20 personnel with limited resources, creating a valid ICE environment.

Fig. 1.

Glacier traverse on James Ross Island.

The scientific activities included a broad spectrum of tasks. In the field of human research, activities comprised the continuous monitoring of physiological functions using wearable sensors, regular testing of cognitive performance, assessment of postural stability, and the implementation of exercise protocols. Environmental research centered on the maintenance of a network of automated meteorological and permafrost stations, glaciological measurements on adjacent glaciers, the collection of geological, hydrological, and biological samples, and the study of periglacial environment dynamics. Furthermore, marine algae were studied as model organisms for life in extreme conditions, and the microbial flora of the station and its crew members was monitored.

Participants

A total of 13 individuals (4 females, 9 males) participated in the postural stability assessment during the mission. One crew member (70 years old) was excluded from the study due to age and slightly reduced mobility. For females, the mean age was 38.8 years (SD = 9.2; range: 28–49), while for males, it was 34.1 years (SD = 7.2; range: 26–48). Mean height was 165.8 cm (SD = 3.6) for females and 179.4 cm (SD = 6.9) for males. Mean weight was 70.0 kg (SD = 3.2; range: 67–74) for females and 79.7 kg (SD = 11.0; range: 52–90) for males. All participants were right-handed and reported no medical conditions, particularly no known neurological, musculoskeletal, or vestibular disorders.

The study was approved by the Institutional Ethics Committee for Biomedical Research of the Faculty of Biomedical Engineering, Czech Technical University in Prague (FBME CTU) (case ID: B11/2025) and was conducted in accordance with international ethical standards. All participants provided written informed consent.

Experimental protocol

The assessment of postural stability was based on a modified Romberg test paradigm¹⁹, which involved measurements under four distinct conditions combining two surface types and two visual feedback states. Participants were instructed to stand barefoot on the force plate, adopting a comfortable, self-selected stance with their arms resting comfortably at their sides. For all trials, they were asked to stand as still as possible. A custom-developed application with a voice assistant guided each participant through the entire measurement sequence, ensuring standardized instructions and timing.

The experimental protocol consisted of four sequential 60-second trials. A fixed order, progressing from easier to more challenging conditions, was used to ensure a consistent experience for all participants and to simplify the self-guided procedure:

1. Firm surface, eyes open (EO): Standing directly on the WBB, looking straight ahead at a designated point.

2. Firm surface, eyes closed (EC): Standing directly on the WBB with eyes closed.

3. Compliant surface, eyes open (EO): Standing on a foam pad (AIREX® Balance-pad Elite, 48 cm40 cm6 cm, AIREX AG, Sins, Switzerland) placed on the WBB, looking straight ahead.

4. Compliant surface, eyes closed (EC): Standing on the foam pad with eyes closed.

Experimental procedure

Due to the complex logistics of the expedition, participants underwent a prolonged multi-leg transit prior to arrival at the station. This included a 10-day stopover in Argentina and a subsequent 9-day stay at Marambio Base. Consequently, upon arrival at the J.G. Mendel Station, the immediate operational priority was station de-winterizing and cargo transport. Therefore, the baseline postural stability measurements commenced at the end of the first week of the station stay.

To ensure data quality in this longitudinal semi-autonomous study, a two-phase approach was implemented:

• Phase 1 (Supervised Training): During the first week of the expedition, all participants performed the protocol under the direct supervision of the investigator. This session served to verify correct foot placement, posture, and adherence to instructions.

• Phase 2 (Unsupervised Monitoring): For the subsequent weeks, participants performed the measurements independently using the self-guided system. To maintain consistency without researcher bias, the custom software provided standardized audio instructions for each trial.

Compliance was verified retrospectively via timestamp logs and visual checks of synchronized RGB video recordings (Intel RealSense D455), which confirmed that participants strictly followed the audio guidance. The retrospective video review was performed for all trials (n = 248) and focused on verifying correct platform placement, adherence to the prescribed sequence of the four measurement conditions, and detecting any stepping off the platform during a trial. No trials were excluded from further analysis based on this review. The purpose of the video inspection was protocol compliance verification rather than the exclusion of trials based on excessive sway or movement characteristics.

Data acquisition and system design

Stabilometric data were acquired using a WBB force platform. A custom, portable system was developed around a Steam Deck OLED (Valve Corp., USA), a handheld console whose capable processor (AMD APU), 16 GB of RAM, and Bluetooth 5.3 connectivity made it an ideal central unit for controlling the experiment and processing data in real-time.

A dedicated application was developed in Python with a PySide6²⁰ graphical user interface. On the device’s touchscreen, participants selected their identity and the experimental protocol. A voice assistant then guided them through the measurement. The assistant used a local text-to-speech engine (pyttsx3²¹) with high-quality audio prompts that were pre-generated by an AI TTS system, ensuring both offline functionality and clarity.

The Steam Deck communicated with the WBB via Bluetooth. During each trial, the stream of data from the four force sensors was continuously recorded. Each resulting comma-separated values (CSV) file contained high-resolution timestamps and the vertical ground reaction force component for each sensor, organized in a hierarchical directory structure (e.g., ‘participant_ID/session_date/trial_type.csv‘). In parallel, to ensure protocol compliance during unsupervised sessions, synchronized RGB video streams were recorded for each trial using the Intel RealSense D455 camera. These recordings were reviewed retrospectively to visually verify participant posture, foot placement, and adherence to the eyes-open/eyes-closed condition.

The entire system was designed for autonomous, self-service use, allowing for standardized measurements without direct researcher supervision. This approach simplified data collection logistics in the isolated station environment and demonstrated the feasibility of using automated diagnostic systems for longitudinal research.

Data quality control and validation

Prior to the main analysis, a detailed data quality control (QC) procedure was conducted to verify the integrity and stability of each measurement’s time series. All QC assessments were performed on a 30-second analysis window, the selection of which is described in section “Time normalization, analysis window, and resampling”.

Analysis of sampling stability

A stable sampling frequency () is critical for reliable time- and frequency-domain analyses. To validate this condition, the instantaneous sampling frequency () was calculated for each measurement from the original high-resolution timestamps. Signal stability was quantified using the following metrics:

• Robust quantification of Jitter: The stability of was quantified using the Relative Median Absolute Deviation (rMAD). This metric was chosen for its robustness against outliers, such as large transmission gaps, which can significantly skew classical measures of dispersion²². A low rMAD value indicates high stability of the sampling frequency. The calculation was performed using Eq. (1):

  1

  where:

  • – Relative Median Absolute Deviation, expressed as a percentage,
  • – instantaneous sampling period (),
  • – median of the set of sampling periods .

• Detection of High-Frequency Bursts: Instances of high-frequency bursts, defined as  Hz, were detected to identify potential transmission errors.

• Detection of Transmission Gaps: Transmission gaps were identified wherever the fell below two critical thresholds. These thresholds were derived from literature recommendations and the requirements of subsequent signal processing steps, ensuring compliance with the Nyquist-Shannon sampling theorem:

  • 50 Hz Threshold: This primary threshold represents the minimum acceptable for clinical stabilometry, based on established recommendations²³.
  • 26 Hz Threshold: A second, more critical threshold identified gaps where  Hz. This value was chosen as twice the 13 Hz cutoff frequency of the Butterworth filter used in subsequent data processing (see section “CoP calculation and filtering”).

Artifact detection and quantification

The second QC step involved the detection and quantification of signal artifacts. Artifacts were defined as rapid, non-physiological changes in the vertical ground reaction force from the individual sensors (Top-Right (TR), Top-Left (TL), Bottom-Right (BR), and Bottom-Left (BL)). They were identified based on the first absolute difference of the signal, calculated using Eq. (2):

2

where is the signal value at time t. The interquartile range () of these differences was then computed. An adaptive threshold, , was set according to Eq. (3):

3

Samples for which were flagged as artifacts.

Data preprocessing

Time normalization, analysis window, and resampling

The original high-resolution timestamps were first converted into a relative time axis, starting at zero for each measurement. For all subsequent analyses, a stable 30-second window²³ from 15 s to 45 s of each 60-second trial was extracted to exclude transient effects during the initial stabilization phase and any anticipatory movements at the end of the trial.

Within this analysis window, the four force signals (TR, TL, BR, BL) were resampled onto an evenly spaced time grid at  Hz ( s) using a shape-preserving piecewise cubic Hermite interpolator (PCHIP²⁴). Prior to interpolation, duplicate or non-increasing timestamps were removed to guarantee a strictly monotonic time axis.

CoP calculation and filtering

The mediolateral () and anteroposterior () coordinates of the Center of Pressure (CoP) were calculated from the raw sensor data using the following standard equations Eqs. (4)–(6):

4

5

6

where:

• – total vertical force,

• – CoP coordinate in the anteroposterior direction,

• – CoP coordinate in the mediolateral direction,

• TR, TL, BR, BL – instantaneous force values from the four sensors,

• d – depth of the force plate (distance between front and rear sensors; 433 mm),

• w – width of the force plate (distance between left and right sensors; 238 mm).

The resulting and time series were then filtered using a 4th-order Butterworth low-pass filter with a cutoff frequency of 13 Hz²⁵. Since the data were processed offline, a zero-phase filter was applied to prevent any phase distortion in the signal.

RM/TR decomposition

The pre-filtered CoP signal was then decomposed into its Rambling (RM) and Trembling (TR) components¹⁸. The slow component (RM) was obtained by applying a 4th-order Butterworth low-pass filter with a 0.5 Hz cutoff frequency¹⁷ to the pre-filtered CoP trajectories. The fast component (TR) was subsequently defined as the residual according to Eq. (7):

7

This frequency-based decomposition was used exclusively for calculating time-domain parameters (i.e., Mean CoP Velocity, Root Mean Square, Sway Area) for the RM and TR components separately. To preserve the original signal’s frequency content, frequency-domain and nonlinear analyses were conducted solely on the full 13-Hz low-pass filtered CoP signal, bypassing the RM/TR decomposition.

Quantification of the CoP trajectory

All parameters calculated from the CoP trajectories for the subsequent statistical analysis are summarized in Table 1.

Table 1.

Summary of the quantified postural control parameters, their symbols, and units.

Parameter	Symbol	Units
Time-domain parameters
Mean Velocity (MV)		mm/s
Sway Area (SA)
Root Mean Square (RMS)		mm
Frequency-domain parameters
Median Freq. (MDF)	,	Hz
Energy Bands (AUC)
Nonlinear parameters
Fractal dimension	,	unitless
DFA exponent	,
Lyapunov exponent	,
MSE complexity index	,
Ratio parameter
Romberg’s quotient		unitless

Note: AP, anteroposterior; ML, mediolateral; total, full signal without decomposition; RM, rambling component of the CoP trajectory; TR, trembling component of the CoP trajectory. The subscript “parameter” for RQ denotes that the quotient was calculated for all previously listed parameters

Time-domain parameters

Mean CoP velocity

Mean CoP Velocity (MV) is a widely used parameter due to its high test-retest reliability across various postural control conditions²⁶. It quantifies the average speed of CoP motion and is calculated from the total path length of the CoP trajectory over the measurement duration, T, as shown in Eq. (8):

8

where:

• MV – Mean CoP Velocity,

• T – total duration of the measurement window in seconds,

• N – total number of samples in the time series,

• – CoP coordinate in the anteroposterior direction at sample i,

• – CoP coordinate in the mediolateral direction at sample i.

Sway area

Sway Area (SA), defined as the area of the 95% confidence ellipse enclosing the CoP data points, is another widely used measure of postural sway²⁷. Its calculation is based on the eigenvalues of the covariance matrix of the CoP coordinates, as shown in Eq. (9):

9

where:

• SA – Sway Area,

• – critical value from the F-distribution for a 95% confidence level with 2 and degrees of freedom,

• N – total number of samples in the time series,

• and – eigenvalues of the covariance matrix of the CoP coordinates.

Root mean square: The Root Mean Square (RMS) distance assesses the variability of CoP displacement²⁸. This parameter is equivalent to the standard deviation of the CoP time series and represents the magnitude of sway around the mean CoP position. RMS was calculated separately for the anteroposterior () and mediolateral () directions. The formula for the AP direction is given in Eq. (10):

10

where:

• is the Root Mean Square value for the anteroposterior direction,

• N is the total number of samples in the time series,

• is the CoP coordinate in the anteroposterior direction at sample i,

• is the mean position of the CoP in the anteroposterior direction.

Frequency-domain analysis

To analyze the frequency characteristics of postural sway, the Power Spectral Density (PSD) was estimated for the and time series using Welch’s method²⁹. A Hann window of 256 samples with a 50% overlap was applied. From the resulting power spectrum, the power within specific frequency bands was quantified by calculating the Area Under the Curve (AUC) via trapezoidal numerical integration. Following the definitions from Fujimoto et al.³⁰, three bands were analyzed: low-frequency (LF: 0.04–0.1 Hz), mid-frequency (MF: 0.1–1.0 Hz), and high-frequency (HF: 1.0–10.0 Hz). Compared to Fujimoto et al.³⁰, the lower cutoff frequency was set to 0.04 Hz instead of 0.02 Hz, taking into account the 30-s analysis window, for which the theoretical minimum resolvable frequency is approximately 0.033 Hz.

Additionally, the Median Frequency (MDF) was computed. The MDF is the frequency that divides the total power of the PSD spectrum into two equal halves. Shifts in the MDF are indicative of changes in neuromuscular control strategy, specifically a shift towards lower frequencies suggests a transition from a Trembling (TR) dominant strategy to a Rambling (RM) dominant one^18,31.

Nonlinear analysis

Fractal dimension

The complexity of the CoP trajectory was assessed using the Fractal Dimension (FD)³². FD was computed using the robust Higuchi algorithm³³, which is well-suited for time-series analysis. This parameter quantifies the space-filling properties of the CoP trajectory, where higher values indicate a more complex and irregular signal, while lower values suggest a simpler, more predictable trajectory.

The calculation was performed separately for the anteroposterior () and mediolateral () axes using the antropy library³⁴. For a given time series, e.g., , k new, shorter time series are first constructed. The length for each of these series is defined by Eq. (11):

11

where m is the starting point index () and k is the time interval. The average length for a given k is then calculated as the mean of all lengths values according to Eq. (12):

12

This average length is expected to scale with the time interval k according to a power law. After log-transforming this relationship, a linear dependency emerges, where the fractal dimension (e.g., ) is defined as the negative of the slope of this relationship, as expressed in Eq. (13):

13

In practice, is obtained as the slope of a least-squares regression line fitted to the data points for a selected range of k values (a standard was used in this study).

Detrended fluctuation analysis

The long-range correlations of the CoP time series were evaluated using Detrended Fluctuation Analysis (DFA)³⁵. DFA quantifies the persistence of patterns in a signal over time. This structure is characterized by the scaling exponent , where a value of indicates an uncorrelated random process (white noise). Values in the range indicate persistence, values in indicate anti-persistence, and suggests a non-stationary, Brownian motion-like process. The calculation was performed separately for the AP () and ML () axes using the nolds library³⁶. For a time series , a cumulative sum profile y(k) is first created according to Eq. (14):

14

This profile is then divided into non-overlapping windows of length n. In each window, a local linear trend is removed, and the root-mean-square fluctuation F(n) of the detrended series is calculated. This fluctuation is expected to scale with the window size n according to a power law shown in Eq. (15):

15

After log-transforming this relationship, the scaling exponent (e.g., ) is determined as the slope of the line relating to . In practice, this is the slope of a least-squares regression line fitted to the data points over a range of window sizes n.

Largest Lyapunov exponent

The local dynamic stability of the postural control system was assessed using the Largest Lyapunov Exponent (LLE)³⁷. A higher LLE value indicates a less stable, more chaotic system, whereas a lower value signifies a more stable and predictable system. The LLE was computed separately for the AP () and ML () axes using the nolds library³⁶. This first requires a state-space reconstruction from the one-dimensional time series (e.g., ) using the method of time-delayed vectors, as shown in Eq. (16):

16

where m is the embedding dimension and is the time delay. For each point on the trajectory , its nearest neighbor is found, and the divergence of their distance d(t) is tracked over time. The LLE is defined as the slope of the linear region of the mean logarithmic divergence curve , as defined by Eq. (17):

17

In this study, an embedding dimension of was used^38–40, and the time delay was determined automatically from the first minimum of the signal’s autocorrelation function³⁷.

Multiscale entropy

The complexity of the time series was evaluated using Multiscale Entropy (MSE)⁴¹. The final metric, a Complexity Index (CI), was computed separately for the AP () and ML () axes using the PyEEG library⁴² and a custom coarse-graining function. The procedure begins by creating 10 coarse-grained time series from the original signal at scale factors , according to Eq. (18):

18

Next, Sample Entropy (SampEn)⁴³ is calculated for each of the 10 coarse-grained series. SampEn is based on comparing template vectors of length m. The entropy value quantifies the conditional probability that two vectors similar for m points will remain similar at the next point, , as defined by Eq. (19):

19

where A is the number of template vector pairs of length having a distance , and B is the number of pairs of length m with a distance . The calculation parameters were: embedding dimension , tolerance radius of the original time series⁴⁴, and the Chebyshev distance ( norm).

Finally, the Complexity Index (CI) is the sum of the SampEn values across all 10 scales, as shown in Eq. (20):

20

A higher CI value is characteristic of a signal with meaningful complexity across multiple time scales, which is considered a feature of a healthy and adaptive postural control system⁴⁴.

Romberg quotient

To quantify the reliance of the postural control system on visual input, the Romberg Quotient (RQ) was calculated for parameters across the time-domain, frequency-domain, and nonlinear analyses⁴⁵. The RQ is defined as the ratio of a given parameter’s value measured with eyes closed (EC) to its value measured with eyes open (EO), as shown in Eq. (21):

21

Statistical analysis

All analyses were conducted using a paired, within-subject design. For each comparison, only complete pairs of measurements, defined by the unique combination of participant ID and measurement date, were included (a complete-case approach). This ensured that all compared conditions originated from the same individuals under the same temporal context.

All pairwise comparisons were conducted using the two-tailed Wilcoxon signed-rank test⁴⁶. To ensure accurate handling of tied ranks and zero differences, the Pratt method⁴⁷ was utilized. The Hodges-Lehmann (HL) estimator⁴⁸ of the median difference () was calculated as a robust measure of effect size.

To account for the large number of comparisons, the Benjamini-Hochberg⁴⁹ procedure was applied to control the False Discovery Rate (FDR) across the entire set of tests performed. Statistical significance was based on the resulting q-values, with a threshold of .

Longitudinal analysis

To address the primary hypothesis regarding changes in postural stability over the course of the expedition, linear mixed-effects models (LMM)⁵⁰ were employed. The models were fitted for each parameter separately, with the measurement week (as a categorical variable) included as a fixed effect. To account for the repeated-measures nature of the data, a random intercept was included for each participant. Degrees of freedom and p-values were estimated using Satterthwaite’s method⁵¹, as implemented with LMM in the lmerTest R package⁵².

Condition contrast analysis

Two types of contrasts were analyzed to assess the influence of sensory conditions: (i) the effect of vision (Eyes Open vs. Eyes Closed), evaluated separately on the firm and compliant surfaces, and (ii) the effect of surface type (firm vs. compliant), evaluated separately under EO and EC conditions. To ensure the stability of the estimates, comparisons with fewer than five complete pairs were not tested and are reported as not applicable (n/a).

Analysis of Romberg quotients

For each parameter, the Romberg Quotient () was first calculated at the individual level, separately for the firm and compliant surfaces. The resulting RQ values were then compared between the two surfaces (compliant vs. firm) using the paired Wilcoxon signed-rank test. During the RQ calculation, any instances of zero in the denominator (the EO value) were treated as missing data (NaN) to prevent division-by-zero errors; these non-finite RQ values were excluded from the analysis.

Presentation of results

The main text illustrates results using selected example parameters (e.g., , , , MDF, , , , and selected RQs), reported as Hodges–Lehmann estimates of the paired median difference () together with their corresponding FDR-adjusted q-values. Comprehensive results for all calculated parameters, including full descriptive statistics (median [IQR]) and unadjusted p-values, are provided in the Supplementary Tables.

Results

A total of 248 postural stability measurements were acquired from 13 subjects over the course of the expedition. The number of measurements per subject varied. Four subjects completed 16 measurements, seven completed 20, and one subject completed 24.

Data quality and preprocessing summary

The data quality control analysis was performed on all 248 measurements, with the results summarized in Table 2.

Table 2.

Summary of data quality metrics.

Data quality metric	Median	[IQR]	Units
Sampling characteristics
Effective	56.81	[53.43 – 56.98]	Hz
Jitter (rMAD)	73.11	[63.99 – 74.70]	%
Signal integrity issues
Severe data gaps ( Hz)	7.06	[6.24 – 7.76]	%
All data gaps ( Hz)	38.19	[37.30 – 39.42]	%
High-freq events ( Hz)	20.96	[20.27 – 21.54]	%
Corrected artifacts	0.31	[0.17 – 0.50]	%

Values represent the median and [IQR] computed across all 248 measurement trials.

Expressed as a percentage of the median sampling period.

The median effective sampling frequency was approximately 57 Hz. However, the signal exhibited considerable jitter, and a substantial portion of the data was affected by transmission gaps where the instantaneous frequency fell below 50 Hz (median 38%). In contrast, the occurrence of severe gaps ( Hz), high-frequency events, and correctable signal artifacts was relatively low.

Longitudinal changes in postural stability

To address our primary hypothesis regarding postural adaptation, linear mixed-effects models (LMM) were used to assess changes in stability parameters across the expedition weeks. Contrary to our hypothesis, no significant longitudinal changes were found. For the representative parameter Mean Velocity (), the results of the LMM are detailed in Table 3.

Table 3.

Linear mixed-effects model results for Mean Velocity () across the expedition weeks.

Predictor	Estimate [mm/s]	SE	df	t-value	p-value
Intercept (Week 1)	75.81	4.06	17.86	18.67
Week 2	2.01	2.54	234.79	0.79	0.430
Week 3	2.19	2.54	234.79	0.86	0.389
Week 4	-0.72	2.54	234.79	-0.28	0.777
Week 5	4.39	2.87	235.95	1.53	0.127
Week 6	-0.08	7.25	237.95	-0.01	0.991

The intercept (baseline performance at Week 1) was estimated at 75.81 mm/s (SE = 4.06, ). Compared to this baseline, there were no statistically significant differences in at Weeks 2–5 (all ). At Week 6, the mean velocity was almost identical to the baseline (estimate = – 0.08 mm/s, ), indicating no trend toward improvement or deterioration over the expedition period.

Similar linear mixed-effects models were run for all primary stabilometric parameters. None of these parameters showed significant changes across the weeks of the expedition (all ), indicating a stable postural control performance throughout the Antarctic mission.

Condition-dependent changes in postural stability

The effect size is expressed as the Hodges–Lehmann estimator of the median difference () and is reported for all comparisons. Statistical significance was based on FDR-corrected q-values, with a threshold of .

Time-domain parameters

The stability of the upright stance, quantified by time-domain parameters, was profoundly influenced by the interplay between visual feedback and surface compliance. This interaction is visually summarized for the key Root Mean Square (RMS) of sway parameters in Fig. 2. The figure clearly illustrates that this sensory-driven instability was substantially more pronounced in the anteroposterior (AP) direction compared to the mediolateral (ML) direction. This directional dependence underscores the critical reliance on vision for controlling AP sway, particularly when proprioceptive cues are unreliable.

Fig. 2.

Anteroposterior and mediolateral postural sway across the four experimental conditions. The boxplots show the Root Mean Square (RMS) of the total Center of Pressure displacement in (A) the anteroposterior (AP) direction (RMS_AP,total) and (B) the mediolateral (ML) direction (RMS_ML,total). Data are aggregated across all measurement sessions throughout the expedition, as no significant longitudinal changes were observed.

Effect of visual feedback

On the firm surface, visual feedback had a subtle yet significant impact on postural stability. The overall sway velocity, measured by , increased slightly ( mm/s, ), a change primarily attributed to a substantial increase in the slower Rambling (RM) component (, mm/s, ). Sway area () also grew ( mm, ), and a significant increase in RMS was observed specifically in the AP direction, while changes in the ML direction were not significant. These findings are detailed in Supplementary Table S1.

When the same test was performed on the compliant surface, the effects were far more pronounced. The loss of vision led to a substantial deterioration in all stability metrics. increased significantly ( mm/s, ), and its slow RM component more than doubled ( mm/s, ), suggesting major, slow corrections were needed to maintain balance. The overall sway area () showed a pronounced increase ( mm, ). A significant increase in RMS was observed in both the AP and ML directions. A full comparison is provided in Supplementary Table S2.

Effect of surface type

The compliant surface consistently impaired postural stability, regardless of visual input. With eyes-open, this was evident across several parameters, including ( mm/s, ), ( mm, ), and ( mm, ). Differences in the ML were not significant. The only minor exception was observed for ( mm, ). A full comparison is provided in Supplementary Table S3.

With eyes-closed, the compliant surface caused an even greater deterioration in performance. increased substantially ( mm/s, ), and showed a pronounced increase ( mm, ). Unlike the eyes-open condition, the ML was also significantly affected, with increasing by mm (). These data are detailed in Supplementary Table S4.

Frequency-domain parameters

Closing the eyes shifted signal power towards mid- and high-frequencies, an effect that was substantially magnified on the compliant surface.

Effect of visual feedback

On the firm surface, closing the eyes led to a significant increase in power in the mid-frequency (, , ) and high-frequency (, , ) bands in the AP direction, and also in the ML direction for the mid-frequency band (, , ). This suggests an increased muscle activity to control small sways. A full comparison is provided in Supplementary Table S5.

The power shift was substantially magnified on the compliant surface. Here, significant increases were evident across all power bands in both the AP and ML axes, indicating that the body shifted its control strategy toward higher-frequency, more active corrections in response to the loss of visual input. A full comparison is provided in Supplementary Table S6.

Effect of surface type

The compliant surface shifted the spectrum of oscillations towards higher frequencies, an effect that was amplified when vision was removed.

With eyes-open, the compliant surface increased power in the mid- and high-frequency bands in the AP direction ( for , ; for , ). Changes in the ML direction were not significant. A full comparison is provided in Supplementary Table S7.

With eyes-closed, this effect was further amplified. A pronounced increase in power was observed in both AP and ML directions, confirming that the compliant surface forces the body to adopt a more aggressive, high-frequency control strategy, particularly when visual cues are absent. A full comparison is provided in Supplementary Table S8.

Nonlinear parameters

Closing the eyes reduced fractal complexity and increased persistence (DFA), with a more pronounced effect on the compliant surface.

Effect of visual feedback

On the firm surface, closing the eyes led to significant changes in the dynamics of CoP movement. The fractal dimension () decreased (, ), suggesting the CoP trajectory became less complex and more predictable. Concurrently, DFA exponent increased (, ), indicating a more persistent, random-walk-like pattern. The Largest Lyapunov Exponent () also increased (, ), signifying a decrease in local dynamic stability. No significant changes were observed for MSE-CI in either direction. For a full comparison, see Supplementary Table S9.

On the compliant surface, these changes were even more pronounced. (, ) and (, ) decreased significantly, while the DFA exponent and LLE increased significantly in both directions. This suggests that without visual feedback, the postural system adopts a less complex, less flexible, and locally more unstable control strategy to maintain balance. The MSE-CI showed a significant increase in the AP direction (, ) and a significant decrease in the ML direction (, ). For a full comparison, see Supplementary Table S10.

Effect of surface type

The compliant surface reduced fractal dimension and increased persistence (DFA), an effect that was magnified with eyes closed.

With eyes-open, standing on the compliant surface significantly reduced the fractal dimension and increased the persistence (DFA) of the CoP signal. This was particularly evident in the AP direction, where decreased (, ) and increased (, ). This indicates that the body shifts from a highly adaptable control strategy to a more rigid one. Changes in LLE were primarily observed in the ML direction (, ), while the multiscale entropy complexity index () indicated a significant reduction in complexity in the AP direction (, ). A full comparison is provided in Supplementary Table S11.

With eyes-closed, the effect of the compliant surface was magnified. decreased with a median difference of (), and decreased by (). DFA increased significantly in both directions (all ). LLE increased significantly only in the ML direction (, ), while changes in the AP direction were not significant. decreased significantly (, ). A full comparison is provided in Supplementary Table S12.

Romberg quotients

Romberg Quotients (RQs) were used to quantify the reliance on vision and showed a clear difference between the two surfaces. On the firm surface, RQs were close to 1, indicating that vision played a minimal role in maintaining stability. However, on the compliant surface, RQs were consistently and significantly higher across time-domain, frequency-domain, and nonlinear parameters. For example, a significant increase in was observed (, ). This pronounced increase confirms that vision becomes the primary source of feedback when proprioceptive cues are unreliable. A full comparison is provided in Supplementary Table S13.

Discussion

This study presents the first comprehensive analysis of postural stability using longitudinal data collected in Antarctica. Our findings provide a detailed characterization of human postural control under the combined stresses of environmental extremes and isolation. We demonstrated that postural stability is profoundly influenced by the interplay of visual and proprioceptive inputs, with nonlinear analyses revealing a strategic shift under sensory stress towards a more rigid control with lower fractal complexity (lower FD), a less corrected, highly persistent dynamic indicating a shift towards Brownian-like motion (higher DFA), and a more reactive, locally unstable state (higher LLE). Contrary to our initial hypothesis, we found no evidence of systematic postural adaptation over the 49-day mission. A key methodological contribution of this study is the successful deployment of a low-cost, highly portable, and autonomous system for field research, which utilized voice guidance to enable standardized, self-administered trials. While this approach proved to be practical and effective for data collection in a remote environment, our detailed data quality analysis also revealed significant technical limitations that require careful consideration. A further limitation is the necessarily small cohort size, which reflects the logistical and safety constraints of Antarctic field missions.

As expected, the sensory manipulation (visual deprivation and compliant surface) elicited significant increases in sway parameters, as evidenced by increases in all time-domain parameters. These results demonstrate the sensitivity and field feasibility of our portable system setup and serve as a manipulation check, aligning with established literature¹⁹. These effects have been well established and, in the present study, primarily serve to demonstrate that our autonomous field system can detect standard postural responses. They are not the main focus of our research question, which concerns longitudinal changes in balance control over the 49-day mission. The interaction between these conditions was particularly strong and is consistent with an increased reliance on vision when proprioceptive feedback is degraded, as described in sensory integration models of postural control^53,54. The significantly higher Romberg Quotient (RQ) values on the compliant surface quantitatively confirm that the central nervous system becomes heavily reliant on visual information when proprioceptive feedback is distorted and unreliable^31,55. Beyond quantifying sway magnitude, our analyses revealed qualitative changes in the underlying control strategy. The concurrent decrease in Fractal Dimension (FD) and increase in the Detrended Fluctuation Analysis (DFA) exponent suggests a shift from a more complex, adaptive control regime to a less complex, more persistent and stereotyped one^32,35. Consistent with this interpretation, spectral analysis showed increased power in the mid- and high-frequency bands, and the RM/TR decomposition indicated a larger Rambling (RM) component^18,30.

From a practical perspective, these findings provide preliminary, mechanistic insights that may inform diagnostic screening and longitudinal monitoring in Antarctic expeditions. Although we did not observe systematic deterioration of postural stability over the 49-day mission, the overall pattern of results replicated well-established effects of visual deprivation and compliant surfaces on postural sway, indicating that our low-cost autonomous system behaves as expected and is sufficiently sensitive to detect standard balance impairments even in an Antarctic ICE setting. In this context, even isolated fall-related injuries can have disproportionate health and operational consequences, as access to advanced medical care is limited and medical evacuation options are constrained¹⁴. Precisely because the sensory manipulations produce the expected group-level response, repeated assessments under these standardized conditions could, in future work, be used to identify individuals whose postural stability deteriorates over time or deviates markedly from the typical response pattern. Such individuals might be at increased risk when operating on uneven or low-friction terrain under reduced visibility and could benefit from closer monitoring or individually tailored preventive strategies. Similar longitudinal screening concepts may be conceptually transferable to other isolated and extreme environments, including long-duration exploration missions, with the important caveat that the present findings apply strictly to terrestrial 1g balance control and not to altered-gravity conditions.

Contrary to our exploratory hypothesis, which presumed a biphasic pattern of initial deterioration followed by recovery, we observed no significant systematic changes in postural stability during the 49-day expedition. This null finding can be attributed to four primary, potentially interacting, factors. (i) A critical methodological limitation was the preclusion of postural assessments during the initial expedition phase due to logistical challenges (specifically, the cumulative fatigue from the multi-leg transit and immediate station de-winterizing), making it impossible to quantify any acute responses to transit stressors as baseline measurements commenced at the end of the first week post-arrival. (ii) The temporal dynamics of adaptation offer two contrasting possibilities: a rapid positive adaptation in the highly fit cohort may have occurred too quickly for our weekly measurement protocol to detect, whereas slower maladaptive neuroplastic processes, which would manifest as a gradual deterioration in stability, may not have had sufficient time to fully develop during the 49-day expedition, as has been observed during a 14-month mission³. (iii) It is plausible that specific neuro-motor adaptations required for navigating the challenging Antarctic terrain during the austral summer (e.g., ice, snow, and exposed rocks) are highly context-specific and may not be detectable by static posturography on a flat surface. (iv) Furthermore, the specialized expedition attire worn outdoors, including multi-layered restrictive clothing and heavy, rigid boots⁵⁶, fundamentally alters the body’s biomechanics and promotes a distinct postural strategy (e.g., hip-dominant control over ankle strategies⁵⁷). While the postural system would undoubtedly adapt to these constraints, such an adaptation would be specific to compensating for the heavy, restrictive gear. This specialized adaptation is maladaptive for the unencumbered, quiet stance tasks performed indoors in light clothing, which rely on sensitive ankle-based control. Therefore, any such changes would not manifest as an improvement in our specific stability metrics and could have been masked. This overall lack of positive transfer could have effectively concealed functional adaptations. It is also important to emphasize that the absence of statistically significant longitudinal changes should not be interpreted as definitive evidence of stability. Rather, it reflects the limited sensitivity of our study design, constrained by sample size, measurement frequency, and mission duration. Longer exposure or larger cohorts may reveal subtle adaptive or maladaptive trends that remained undetectable in the present dataset.

A key methodological finding of this study is the successful field deployment of the Wii Balance Board (WBB) within a low-cost, portable, and autonomous data collection system. The use of a self-guided, voice-narrated protocol proved highly effective, enabling robust data collection without researcher supervision in an Antarctic environment. Rather than providing a de novo validation of the instrument, our findings demonstrate the feasibility and field performance of this approach in an ICE setting, building on previous validation studies of the WBB conducted under controlled conditions^15,16. However, our detailed data quality analysis revealed significant technical limitations that demand careful consideration for future applications. The WBB signal was characterized by high jitter and frequent transmission gaps, posing considerable challenges for processing. These signal integrity issues are likely inherent to the WBB’s Bluetooth protocol but were potentially exacerbated by the specific communication interface with the non-standard hardware platform (Steam Deck). Critically, our analysis showed that in some data segments (median 7%), the instantaneous sampling frequency dropped below 26 Hz. This violates the Nyquist-Shannon sampling theorem for our 13 Hz low-pass filtered signal, potentially introducing aliasing artifacts and compromising the reliability of frequency-domain parameters. Furthermore, such temporal instability can distort the estimation of nonlinear parameters, which are highly sensitive to the temporal structure of the time series. Therefore, while the WBB system is a viable and practical tool for collecting time-domain parameters in challenging field settings, our findings serve as a critical cautionary note. Beyond the Antarctic context, the feasibility of a self-guided, voice-narrated stabilometry protocol highlights the potential for broader applications. Similar low-cost, portable, and autonomous systems could be valuable in telemedicine, rehabilitation, or geriatric fall-prevention programs, where frequent standardized assessments are needed outside of laboratory or clinical environments.

The principal limitations of this study include the small sample size, the fixed order of experimental conditions necessitated by the self-guided protocol, technical constraints associated with the WBB data acquisition system, and the differences between the indoor testing environment and the actual biomechanical challenges of the Antarctic terrain. Because the measurements were performed using a self-guided system without direct supervision, the progression from easier to more challenging conditions was deliberately kept fixed across subjects and measurement sessions to minimize procedural errors. However, this non-randomized order may have introduced order effects (e.g., warm-up or learning), potentially leading to a conservative underestimation of impairments in the later, more challenging conditions. The restricted sample size (n = 13) did not result from recruitment bias but was determined by the expedition architecture. The mission involved 14 individuals, which represented the logistical and safety limit of the station’s capacity, and one member (70 years, with slightly reduced mobility) was excluded a priori based on predefined eligibility criteria. This inherent ceiling inevitably reduces statistical power and limits the generalizability of the findings, particularly for subgroup analyses and nonlinear metrics. Nevertheless, it provides an authentic representation of the constraints under which research in isolated and capacity-limited environments must be conducted. Moreover, we did not estimate classical test-retest reliability indices (e.g., intraclass correlation coefficients) for the WBB in this Antarctic field setting; our assessment was limited to feasibility and signal quality metrics rather than formal reliability testing. Finally, due to the intense operational schedule and competing demands from the multidisciplinary research program, a formal post-expedition debriefing regarding subjective protocol adherence was not conducted. However, as noted in the Methods, this limitation was mitigated by the objective retrospective analysis of video recordings. Specifically, inspection of all video recordings confirmed full protocol adherence, and no trials were excluded.

Future research should aim to address these limitations by utilizing high-fidelity portable force platforms, performing measurements in low-temperature environments and in full outdoor gear. The integration of electromyography (EMG), inertial measurement units (IMUs), and instrumented pressure insoles is also recommended. While EMG and IMUs provide insight into neuromuscular activity and body kinematics, pressure insoles capture plantar pressure distribution and allow the estimation of foot-ground interaction and center of pressure trajectories even during locomotion in full expedition footwear. Together, these complementary modalities would enable a more comprehensive assessment of neuro-motor control strategies during both quiet stance and real-world tasks, thereby improving our understanding of adaptation in these challenging environments.

Author contributions

P.V.: Conceptualization, Methodology, Formal Analysis, Funding acquisition, Project administration, Supervision, Investigation, Writing – Original Draft. M.S.: Formal Analysis, Data Curation, Software, Writing – Original Draft, L.L.: Investigation, Visualization, Writing – Review & Editing. J.H.: Investigation, Validation, Writing – Review & Editing. P.K.: Conceptualization, Methodology, Supervision, Validation, Writing – Review & Editing. All authors reviewed the manuscript.

Funding

This work was supported by the Technology Agency of the Czech Republic, grant TM05000017 ”Portable modular MoCap system for recording and evaluating the condition of members of the armed forces and first responders”.

Data availability

The datasets generated and analyzed during the study are not publicly available due to ethical considerations and participant protection, but they are available from the corresponding author upon reasonable request.

Competing interests

The authors declare no competing interests.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-026-36215-9.