BACKGROUND:Human Activity Recognition (HAR) and exercise assessment models are increasingly used in healthcare to support clinical evaluation, rehabilitation, and remote monitoring. However, their real-world applicability critically depends on the ability to generalize across unseen subjects, whose movement patterns may differ substantially due to inter-individual variability. Despite this, many studies adopt random noncross-subject (NCS) data splits, where samples from the same individual appear in both training and test sets, potentially leading to overly optimistic and clinically misleading performance estimates. OBJECTIVE:We investigate (i) how NCS and cross-subject (CS) splits affect performance estimation across machine learning and deep learning models under tasks of increasing complexity, (ii) how data splitting and differences between training and test sets contribute to predictive variance and stability. METHODS:Experiments were performed using a large-scale HAR benchmark dataset (NTU RGB+D 120) and a rehabilitation-specific dataset (IntelliRehabDS). A total of 12 machine learning and deep learning models were trained across both tasks, and their performance was estimated and compared using a simulation-based approach. Predictive variance decomposition, via Generalized Linear Mixed-Effects models, was applied to link the split strategy and differences in training and test instances to model output stability. RESULTS:NCS splits consistently overestimated model performances, with discrepancies increasing alongside task and model complexity. DL architectures, in particular, showed markedly higher NCS performance compared to CS splits, generally with statistical significance. Variance decomposition revealed that greater subject difference between training and test sets often enhances predictive instability, while CS splitting reduces variance by promoting more generalizable representations. CONCLUSIONS:Improper dataset splits can mislead model evaluation, exaggerate generalization capabilities, and undermine clinical trust. Our study provides empirical evidence for computer vision-based rehabilitation models and offers methodological guidance for robust evaluation practices, supporting reproducible and trustworthy AI deployment in rehabilitation and broader healthcare applications.

Object detection, a cornerstone of computer vision powered by advancements in Convolutional Neural Networks (CNNs), plays a crucial role in enabling robots to perceive and interact with their surroundings, particularly in complex applications such as rehabilitation robotics. This paper investigates the viability of integrating the real-time-capable YOLOv11 architecture into the TIAGo robot for object detection tasks relevant to rehabilitation settings. Given the limitations of TIAGo's LiDAR—especially its fixed height, which hinders obstacle detection—this study explores whether YOLOv11 applied to RGB data from the robot's onboard camera can compensate for such perceptual gaps. We apply transfer learning to fine-tune various YOLOv11 variants (n, s, m, l, x) using a publicly available dataset of indoor scenes acquired with RGB-D camera, sensor frequently on board on assistive robot. Each model is evaluated in terms of detection accuracy (mAP50), inference time, and memory usage to assess its suitability for deployment under real-time constraints imposed by TIAGo's 30 Hz RGB-D camera. Considering the technical specifications of TIAGo, our results show that YOLOv11-s achieves the highest mAP50 (96.4%) but exceeds the frame rate requirement with an inference time of 34.6 ms, suggesting that optimization would be necessary for robotic purpose. The analysis highlights trade-offs between accuracy and computational efficiency and supports the feasibility of future integration of object detection within TIAGo's navigation framework for safe and effective rehabilitation assistance.

Exoskeletons are a fast-growing technology that enables multiple use-cases in clinical scenarios. They can be useful tools for the rehabilitation of patients with motor dysfunctions caused by neurological conditions, aging or trauma. Assistive exoskeletons modulate the torque exerted by the electrical motors moving their joints to allow the patients wearing them to achieve an intended movement, such as gait, correctly. Their effectiveness, therefore, requires accurate online control of such torques to complement those generated by the patient. Hereby we explored Deep Learning (DL) models to generate an online prediction of the gait phase, i.e., stance or swing, during assisted walking with a lower-limb exoskeleton based on surface electromyography (sEMG) data. We leveraged the lead of muscular activation with respect to the movement of the limbs to adjust the labeling based on joints kinematics. The cross-subject design allowed to generalize over subjects not considered for training A hyperparameter optimization algorithm was also implemented to further explore the capabilities of DL models of a reduced size. We simulated a use case scenario to assess whether online implementation of the proposed technique is feasible. We also proposed a new metric called trade-of score (TOS) for evaluating the cost-performance compromise of the optimized models which lead to identifying a DL model capable of classifying gait phases with an accuracy of about 95% while significantly reducing the number of parameters compared to the full architecture. Its mean computational time of less than 10 ms offers the opportunity for accurate, online exoskeleton control based on sEMG data.

Hemispatial inattention (HI) is a disabling and often persistent consequence of right-hemisphere stroke, typically resistant to conventional rehabilitation approaches. This study reports the results of preliminary tests on a modular robotic platform developed to support HI rehabilitation by integrating autonomous navigation, patient identification, speech-based interaction, patient response assessment, and adaptive therapy delivery. The system leverages multimodal sensing, including RGB-D cameras, skeletal tracking via the MediaPipe framework, and automatic speech recognition, to deliver personalized sensorimotor and cognitive exercises targeting the neglected hemispace. The navigation protocol allows the correct positioning of the robot in front of the patient. The monitoring of the session leverages markerless head pose estimation to study the patient response to the delivered stimuli in terms of spatial exploration. Verbal input is interpreted through a rule-based intent recognition module, enabling context-aware dialogue management and real-time adaptation of the therapeutic session. A preliminary feasibility study was conducted in the Living Lab of the CoE REDI (Centre of Excellence on Rehabilitation Devices and Digital Instruments) at the University of Pavia. The results demonstrated the platform's ability to accurately identify multiple individuals simultaneously, correctly assigning their names and roles. It also achieved a 94% accuracy rate in interpreting spoken commands. Additionally, the system was able to assess patient engagement by leveraging head pose estimation, with a root mean square error (RMSE) of less than 10 degrees. These preliminary results support the potential of the system to deliver structured, responsive, and data-driven rehabilitation experiences. The ongoing work includes full-system integration, deployment of a centralized patient database, and clinical trials to evaluate efficacy and usability in real-world settings.

The Fit4MedRob initiative (“Fit for Medical Robotics”) is a large-scale, multi-partner project funded by the Italian Ministry of University and Research under the National Recovery and Resilience Plan (NRRP). It involves 25 partners, including research institutions, clinical centers, and industrial companies, with the goal of integrating advanced robotic technologies into rehabilitation and assistive care. Within this initiative, Mission 2 - Activity 5 (Match Making #53) specifically targets the development and validation of a robotic protocol for the rehabilitation of post-stroke hemispatial inattention (HI), a disabling condition characterized by the inability to perceive and/or attend to one side of the peripersonal space. This paper presents the design and preliminary evaluation of a modular robotic system based on the TIAGo platform, a collaborative mobile manipulator developed by PAL Robotics. The architecture integrates five key modules: navigation, patient identification, interaction, therapy delivery, and response evaluation. A series of preliminary tests were conducted with healthy participants to validate core components of the system, including autonomous navigation accuracy, real-time face recognition based on ROS4HRI (Robot Operating System for Human-Robot Interaction), and head pose estimation exploiting the MediaPipe library, validated against a ground truth provided by stereophotogrammetric measures. Results confirm the feasibility of the proposed system and support its future application in clinical trials.

Objectives: We assessed the difference between quiet stance and gait in the spatial distribution and intensity of foot plantar pressures and whether it is possible to estimate the distribution during gait from data obtained during stance. Methods: A total of 60 healthy subjects with a mean age of 31.0 ± 9.4 years performed two trials for quiet stance and four trials for gait on a baropodometric walkway with their eyes open. Foot plantar pressures were recorded from 10 areas of the foot sole. Results: During quiet stance, the highest plantar pressure occurred at metatarsal heads (M2 to M4) and the medial (MH) and lateral halves of the heel (LH). During gait, the profile of plantar pressure values was like that during stance, but significantly higher. The differences concentrated at the big toe (T1), M2 to M4, MH, and LH, whilst toes (T2,3,4,5) and midfoot (MF) showed the smallest difference. A significant positive correlation was found between the corresponding areas of foot pressure during gait and stance. Conclusions: During quiet stance and gait, the overall profile of plantar pressure distribution was similar. During quiet stance, the subjects loaded more on the heels, in keeping with the known position of the center of pressure just in front of the ankles. During gait, higher pressures on the metatarsal areas are related to the forward propulsion of the center of mass. The correlation between the corresponding areas of foot pressure during gait and stance suggests that the pressure distribution during gait can partly be estimated from that during stance. This finding might be useful in most clinical settings when a single sensorized platform rather than a complete walkway is available.

Human Activity Recognition plays a crucial role in Ambient Assisted Living, where environmental sensors, including RGB-D cameras like Kinect, are increasingly used due to the many advantages they offer. However, challenges remain in handling occlusions, self-occlusions, and the limited field of view of the device, which can compromise tracking quality, especially when the subject is in motion or in non-optimal positions. In this work, we propose a novel method for combining data from two synchronized and calibrated Kinect devices to address occlusion and self-occlusion issues, improving tracking and expanding the tracking area. The fusion algorithm is based on the confidence level attributed to the joint's coordinates, the orientation and the distance of the subject from the devices. It was then tested on an acquisition where a subject performed a series of daily activities and simulated a dangerous situation such as falling to the floor. Experimental results demonstrate the effectiveness of fusion algorithm in reducing frame loss and improving the quality of the tracking process, particularly when the subject is between 0.5 and 3 meters from the devices and in a standing or sitting pose. However, when the subject is lying down or positioned beyond the optimal range, the method's effectiveness decreases.

Balance impairments can affect walking ability and, consequently, limit physical activity level (PAL). This investigation in facioscapulohumeral muscular dystrophy (FSHD) may help unravel factors limiting PAL and identify new rehabilitation strategies to improve mobility in these patients. Eighteen FSHD patients (10 males; mean age: 36.33 years) and twenty-three age- and sex-matched healthy controls (HC) were recruited. Balance was assessed using the Mini-BEST and Four Step Square test (FSST), while PAL was measured with the International Physical Activity Questionnaire. Gait performance during usual walking (UW), fast walking (FW), and dual-task walking (DT) was evaluated using an inertial sensor (BTS G-Walk). Gait speed (GS), stance time (ST), gait cycle (GC), stride length (SL), pitch and roll angles were considered. Correlations between gait parameters and Mini-BEST were performed in patients, while between PAL, GS, and SL were assessed in both groups. Compared to HC, patients showed lower PAL, Mini-BEST scores and longer FSST times. As group effect, patients exhibited major ST, GC, and roll angle. In the interaction between group and walking condition, during FW patients showed reduced GS and prolonged GC, while no differences were observed in UW and DT. Altered gait parameters, except roll angle, were associated with Mini-BEST, while PAL correlated with gait parameters only in HC. Balance resulted associated with impaired gait parameters in FSHD. The association between balance and gait deficits emerging during FW indicate balance impairment as an element discouraging patients from engaging in physically demanding tasks, plausibly decreasing PAL.

Introduction. Parkinson's disease (PD) can impair both motor and respiratory functions (MFs and RFs), reducing physical efficiency and quality of life. While various exercise modalities have shown benefits on either MFs or RFs, it is unclear whether a multi-component training program (McTP), combining different exercise types, can improve both concurrently. Methods. Ten PD patients (age: 75.1 +/- 7.1 years; 9 males, 1 female), all at Hoehn Er Yahr stage 2, completed a 3-month biweekly McTP. Each 1-hour session included 20 minutes of aerobic exercise (AE) at 14-17 on the Borg scale (targeting RFs), followed by lower-limb resistance and balance exercises (both at Borg level 15) to improve MFs. Training parameters were based on literature in PD or, when unavailable, age-matched healthy cohorts. RFs were assessed via spirometry, 12-second forced inspiration/ expiration, and maximal inspirations. MFs were evaluated with the 6-Minute Walking Test (6MWT), Timed Up Er Go (TUG), and Short Physical Performance Battery (SPPB). Pre/ post comparisons used Wilcoxon signed-rank and paired-tests. Results. No significant changes (p>0.05) were observed in RFs or TUG. However, 6MWT (p<0.01) and SPPB (p <= 0.05) scores improved significantly. Conclusions. These findings suggest that McTP effectively improved MFs, but no statistically significant changes were observed in RFs. Modifying AE frequency and/or duration may be necessary to target respiratory adaptation.

The effect of simultaneously performing two tasks (dual-task effects, DTEs) has been extensively studied, mainly focusing on the combination of cognitive and motor tasks. Given their potentially detrimental impact on real-life activities, the impact of DTEs has been investigated in both healthy individuals and patients. In this Registered Report, we aimed to replicate previous DTEs when a task requiring executive-inhibitory skills is involved while also expanding the evidence on basic facets of decision-making. We recruited 50 healthy young participants who performed a stop-signal task and two gambling tasks (loss-aversion and risk-aversion) while sitting and while walking at three treadmill speeds (normal, slow and fast). We report a significant difference in performance during single-task and dual-task, although with high individual variability. The data show no effect of the walking speed on all the cognitive tasks. Analyses on postural alignments, assessed in the cadence, gait cycle length and stance phase, confirm previous results on cognitive prioritization strategies of healthy individuals. Based on our results, we highlight the need to further investigate prioritization strategies when tasks involving higher cognitive functions are performed along a motor task in healthy individuals and patients with the aim of offering targeted training and rehabilitation protocols. The stage 1 protocol for this Registered Report was accepted in principle on 28/06/22. The protocol, as accepted by the journal, can be found at: https://doi.org/10.17605/OSF.IO/5MWH7 .

Ambient Assisted Living is crucial for improving independence, safety and quality of life in frail individuals. Human Activity Recognition (HAR) can be a game-changing tool to monitor daily life habits, and the progress of rehabilitation that frail individuals undergo. Here we investigated whether deep learning (DL) networks are suitable to identify and quantify the activities that a person performs. We used two Azure Kinect cameras, which enable 3D body tracking while assuring the privacy of the person recorded to acquire a dedicated dataset. The acquisition of the dataset involved twenty healthy subjects performing selected tasks in a set of mixed items including daily living activities and rehabilitation-related tasks. The final dataset was composed of 299 acquisitions of 30 Hz skeletal data. A second version of the dataset was obtained by subsampling the data at 10 Hz to allow more time for processing individual frames. We then implemented three different DL networks: a Bidirectional Long Short-Term Memory (BILSTM), a Temporal Convolutional Network (TCN) and an attention based one for HAR. Finally, we proposed an algorithm processing networks' output to quantify the identified activities in terms of the number of repetitions and time spent performing them. The best network achieves 94% accuracy in HAR and 81% in repetition counting.

Initially developed for entertainment, the Kinect, has become widely used in robotics, rehabilitation, and human-computer interaction thanks to its RGB-D image capture and body tracking capabilities. However, professional applications reveal limitations such as quantization noise, low resolution, and occlusion, which can compromise accuracy. A potential solution is the use of a properly calibrated Kinect network to ensure consistent output across cameras. This study proposes a flexible extrinsic calibration procedure for Azure Kinect cameras, based on 2-D checkerboard corner detection. This method allows independent calibration of each camera without requiring a main-secondary configuration. Calibration was tested with two- and four-camera setups, positioned at heights of 1.2 and 2 m, with varying distances (1-2.29 m) and orientations relative to the checkerboard. Results showed high accuracy, with mean calibration errors (CEs) below 1 cm under optimal conditions, reaching a minimum of 3.3 mm when the checkerboard was placed 1.75 m from the cameras. The proposed method proved effective in multicamera setups, providing large calibration volumes. This study highlights the feasibility and flexibility of the proposed approach, with potential for future research with different sensors and varying conditions.

Multiple Sclerosis (MS) is a chronic neurological disorder affecting the central nervous system that significantly impairs postural control and functional abilities. Robotic-assisted gait training mitigates this functional deterioration. This preliminary study aims to investigate the effects of a four-week gait training with the ExoAtlet II exoskeleton on static balance control and functional mobility in five individuals with MS (Expanded Disability Status Scale ≤ 2.5). Before and after the training, they were assessed in quiet standing under Eyes Open (EO) and Eyes Closed (EC) conditions and with the Timed Up and Go (TUG) test. Center of Pressure (CoP) Sway Area, Antero–Posterior (AP) and Medio–Lateral (ML) CoP displacement, Stay Time, and Total Instability Duration were computed. TUG test Total Duration, sit-to-stand, stand-to-sit, and linear walking phase duration were analyzed. To establish target reference values for rehabilitation advancement, the same evaluations were performed on a matched healthy cohort. After the training, an improvement in static balance with EO was observed towards HS values (reduced Sway Area, AP and ML CoP displacement, and Total Instability Duration and increased Stay Time). Enhancements under EC condition were less marked. TUG test performance improved, particularly in the stand-to-sit phase. These preliminary findings suggest functional benefits of exoskeleton gait training for individuals with MS.

Lower limb exoskeletons represent a relevant tool for rehabilitating gait in patients with lower limb movement disorders. Partial assistance exoskeletons adaptively provide the joint torque needed, on top of that produced by the patient, for a correct and stable gait, helping the patient to recover an autonomous gait. Thus, the device needs to identify the different phases of the gait cycle to produce precisely timed commands that drive its joint motors appropriately. In this study, EMG signals have been used for gait phase detection considering that EMG activations lead limb kinematics by at least 120 ms. We propose a deep learning model based on bidirectional LSTM to identify stance and swing gait phases from EMG data. We built a dataset of EMG signals recorded at 1500 Hz from four muscles from the dominant leg in a population of 26 healthy subjects walking overground (WO) and walking on a treadmill (WT) using a lower limb exoskeleton. The data were labeled with the corresponding stance or swing gait phase based on limb kinematics provided by inertial motion sensors. The model was studied in three different scenarios, and we explored its generalization abilities and evaluated its applicability to the online processing of EMG data. The training was always conducted on 500-sample sequences from WO recordings of 23 subjects. Testing always involved WO and WT sequences from the remaining three subjects. First, the model was trained and tested on 500 Hz EMG data, obtaining an overall accuracy on the WO and WT test datasets of 92.43% and 91.16%, respectively. The simulation of online operation required 127 ms to preprocess and classify one sequence. Second, the trained model was evaluated against a test set built on 1500 Hz EMG data. The accuracies were lower, yet the processing times were 11 ms faster. Third, we partially retrained the model on a subset of the 1500 Hz training dataset, achieving 87.17% and 89.64% accuracy on the 1500 Hz WO and WT test sets, respectively. Overall, the proposed deep learning model appears to be a valuable candidate for entering the control pipeline of a lower limb rehabilitation exoskeleton in terms of both the achieved accuracy and processing times.

Human Activity Recognition (HAR) is pivotal for automating the detection and classification of human movements. Technological advances in data collection from sensors and cameras is making automatic HAR increasingly available to support robotics-assisted rehabilitation, where the detection of abnormal patterns in movement and gait can aid clinical evaluation of patients with diverse disorders. HAR typically employs Machine Learning (ML) or Deep Learning (DL) techniques to analyze sensor data patterns for various activities. However, the generalizability of ML/DL models remains underexplored, particularly concerning the impact of different training and evaluation settings. Overfitting is a common challenge, especially in medical applications where user-dependent training data hinder generalization. Investigating the performance of ML and DL algorithms after a cross-subject (CS) and a noncross-subject (NCS) division of the dataset into training and test sets is crucial. This study aims to systematically compare various ML and DL algorithms for HAR, evaluating their performance across the CS and NCS settings to discern the impact of inter-subject variability on model generalizability.

Walking involves multiple gait cycles comprising stance and swing phases, crucial for stability and propulsion. Recovery of walking is a primary goal in neurological rehabilitation, with robotic exoskeletons emerging as promising tools. These devices assist individuals with lower extremity weakness, utilizing various signal acquisition technologies like inertial sensors (IMUs) to interpret human body movements. Accurately estimating gait phases, particularly for online exoskeleton control, presents a significant challenge. Deep Learning (DL) algorithms offer solutions by analysing IMU signals to predict gait events. A proposed DL model, based on a hybrid solution of one-dimensional Convolutional Neural Network and Bidirectional Long Short-Term Memory Neural Network, utilizes kinematic data from hip, knee and ankle to identify gait phases during exoskeleton-assisted walking, achieving 98.64% accuracy in stance and swing identification. Moreover, the study evaluates the time taken for the prediction of the gait phases, crucial for effective exoskeleton control, estimating the time spent on acquisition, elaboration, and classification procedures. The results show a processing time of 57 milliseconds for a 1-second segment of IMU signal. Evaluated on 26 healthy subjects, the model demonstrates potential for online gait phase identification. Overall, the proposed DL model shows promise for improving exoskeleton-assisted rehabilitation, offering tailored support, and enhancing the effectiveness of gait training interventions.

Exoskeletons are emerging as powerful tools for lower-limb rehabilitation, providing assistance and gait training capabilities. This study investigates the integration of electromyography (EMG) signals with Deep Learning (DL) techniques to identify gait phases (stance and swing) during exoskeleton-assisted walking. A Bidirectional Long Short-Term Memory is trained on walking overground EMG data collected from 26 healthy subjects and tested on both overground and treadmill walking data. The model's online prediction ability is evaluated using as input the Raw (R) data and two different data pre-processing methods: Root Mean Square Value (RMSV) and Savitzky-Golay (SG) smoothing. The model demonstrates adaptability across different walking conditions and achieves precise identification of stance and swing phase within gait data, with the best overall accuracy of about 90% for walking overground RMSV data and about 92% with walking on treadmill SG data. The study also evaluates the time taken for the prediction of the gait phases, crucial for effective exoskeleton control, estimating the time spent for the pre-processing and the classification procedures. The results are of 69.5 milliseconds for R data, 94 milliseconds for SG data and 101 milliseconds for RMSV data, considering a 1-second segment of EMG signal. The SG filter emerges as the preferred methodology as it appears to represent a trade-off between accuracy and processing time. The proposed DL model shows potential for enhancing the precision and responsiveness of exoskeleton-assisted rehabilitation, providing personalized support to individuals undergoing gait training.

The ground reaction force (GRF) recorded by a platform when a person stands upright lies at the interface between the neural networks controlling stance and the body sway deduced from centre of pressure (CoP) displacement. It can be decomposed into vertical (VGRF) and horizontal (HGRF) vectors. Few studies have addressed the modulation of the GRFs by the sensory conditions and their relationship with body sway. We reconsidered the features of the GRFs oscillations in healthy young subjects (n = 24) standing for 90 s, with the aim of characterising the possible effects of vision, support surface and adaptation to repeated trials, and the correspondence between HGRF and CoP time-series. We compared the frequency spectra of these variables with eyes open or closed on solid support surface (EOS, ECS) and on foam (EOF, ECF). All stance trials were repeated in a sequence of eight. Conditions were randomised across different days. The oscillations of the VGRF, HGRF and CoP differed between each other, as per the dominant frequency of their spectra (around 4 Hz, 0.8 Hz and <0.4 Hz, respectively) featuring a low-pass filter effect from VGRF to HGRF to CoP. GRF frequencies hardly changed as a function of the experimental conditions, including adaptation. CoP frequencies diminished to <0.2 Hz when vision was available on hard support surface. Amplitudes of both GRFs and CoP oscillations decreased in the order ECF > EOF > ECS ≈ EOS. Adaptation had no effect except in ECF condition. Specific rhythms of the GRFs do not transfer to the CoP frequency, whereas the magnitude of the forces acting on the ground ultimately determines body sway. The discrepancies in the time-series of the HGRF and CoP oscillations confirm that the body's oscillation mode cannot be dictated by the inverted pendulum model in any experimental conditions. The findings emphasise the robustness of the VGRF "postural rhythm" and its correspondence with the cortical theta rhythm, shed new insight on current principles of balance control and on understanding of upright stance in healthy and elderly people as well as on injury prevention and rehabilitation.

Different measurements of body oscillations in the time or frequency domain are being employed as markers of gait and balance abnormalities. This study investigates basic relationships within and between geometric and spectral measures in a population of young adult subjects. Twenty healthy subjects stood with parallel feet on a force platform with and without a foam pad. Adaptation effects to prolonged stance were assessed by comparing the first and last of a series of eight successive trials. Centre of Foot Pressure (CoP) excursions were recorded with Eyes Closed (EC) and Open (EO) for 90s. Geometric measures (Sway Area, Path Length), standard deviation (SD) of the excursions, and spectral measure (mean power Spectrum Level and Median Frequency), along the medio-lateral (ML) and antero-posterior (AP) direction were computed. Sway Area was more strongly associated than Path Length with CoP SD and, consequently, with mean Spectrum Level for both ML and AP, and both visual and surface conditions. The squared-SD directly specified the mean power Spectrum Level of CoP excursions (ML and AP) in all conditions. Median Frequency was hardly related to Spectrum Level. Adaptation had a confounding effect, whereby equal values of Sway Area, Path Length, and Spectrum Level corresponded to different Median Frequency values. Mean Spectrum Level and SDs of the time series of CoP ML and AP excursions convey the same meaning and bear an acceptable correspondence with Sway Area values. Shifts in Median Frequency values represent important indications of neuromuscular control of stance and of the effects of vision, support conditions, and adaptation. The Romberg Quotient EC/EO for a given variable is contingent on the compliance of the base of support and adaptation, and different between Sway Area and Path Length, but similar between Sway Area and Spectrum Level (AP and ML). These measures must be taken with caution in clinical studies, and considered together in order to get a reliable indication of overall body sway, of modifications by sensory and standing condition, and of changes with ageing, medical conditions and rehabilitation treatment. However, distinct measures shed light on the discrete mechanisms and complex processes underpinning the maintenance of stance.

When a person stands upright quietly, the position of the Centre of Mass (CoM), the vertical force acting on the ground and the geometrical configuration of body segments is accurately controlled around to the direction of gravity by multiple feedback mechanisms and by integrative brain centres that coordinate multi-joint movements. This is not always easy and the postural muscles continuously produce appropriate torques, recorded as ground reaction force by a force platform. We studied 23 young adults during a 90 s period, standing at ease on a hard (Solid) and on a compliant support (Foam) with eyes open (EO) and with eyes closed (EC), focusing on the vertical component of the ground reaction force (VGRF). Analysis of VGRF time series gave the amplitude of their rhythmic oscillations (the root mean square, RMS) and of their frequency spectrum. Sway Area and Path Length of the Centre of Pressure (CoP) were also calculated. VGRF RMS (as well as CoP sway measures) increased in the order EO Solid ≈ EC Solid < EO Foam < EC Foam. The VGRF frequency spectra featured prevailing frequencies around 4–5 Hz under all tested conditions, slightly higher on Solid than Foam support. Around that value, the VGRF frequencies varied in a larger range on hard than on compliant support. Sway Area and Path Length were inversely related to the prevailing VGRF frequency. Vision compared to no-vision decreased Sway Area and Path Length and VGRF RMS on Foam support. However, no significant effect of vision was found on VGRF mean frequency for either base of support condition. A description of the VGRF, at the interface between balance control mechanisms and sway of the CoP, can contribute information on how upright balance is maintained. Analysis of the frequency pattern of VGRF oscillations and its role in the maintenance of upright stance should complement the traditional measures of CoP excursions in the horizontal plane.