Multimodal assessment of balance dysfunction in older adults: from classification to clustering-based functional pattern identification

Abstract

Background

Balance dysfunction in older adults is a major risk factor for falls. Conventional subjective scales and single-modality measures may not adequately capture the integrated motor, cortical, and cognitive processes underlying postural control. Using a multimodal feature set, this study aimed to develop an interpretable framework for dysfunction-related classification and for exploring heterogeneity in balance-related functional patterns.

Methods

A total of 81 community-dwelling older adults (≥ 60 years old) were recruited and divided into balance dysfunction group (n = 38, BBS ≤ 45) and healthy control group (n = 43, BBS > 45) according to the Berg Balance Scale (BBS). During six stance tasks, center of pressure (COP) trajectories, lower-limb surface electromyography (sEMG), and cortical activation measured by functional near-infrared spectroscopy (fNIRS) were synchronously recorded. Group × task effects were tested using two-way repeated-measures ANOVA. Multimodal features were then used for supervised classification (XGBoost) and for clustering-based exploration of heterogeneity within dysfunction-related multimodal profiles using K-means clustering.

Results

Under more challenging postural conditions, the dysfunction group showed larger anteroposterior COP oscillation range, velocity, and area. Muscle activation patterns showed lower rectus femoris contribution together with higher biceps femoris and gastrocnemius activation under selected task conditions, accompanied by altered muscle synergy organization. fNIRS revealed greater activation in premotor cortex, supplementary motor area, and prefrontal cortex regions in the dysfunction group. In classification, XGBoost achieved the best overall performance among the tested models, with 83.56% accuracy. Clustering analysis further identified three functional patterns with graded differences in BBS and Montreal Cognitive Assessment (MoCA) scores, suggesting heterogeneity in dysfunction-related multimodal profiles.

Conclusion

Balance dysfunction in older adults is associated with a neuro-muscular-cognitive profile involving impaired postural control, altered lower-limb muscle recruitment, and increased cortical activation under more demanding task conditions. The proposed multimodal framework provides an interpretable approach for classification and for exploring heterogeneity in dysfunction-related functional patterns. Further validation in larger cohorts and simplified sensor configurations will be needed to support broader practical application.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s12984-026-02095-3.

Background

Balance dysfunction is a clinically relevant functional problem in older adults and is closely associated with fall risk [1, 2]. Previous evidence indicates that balance impairment contributes to an increased risk of falls in community-dwelling older adults, while international fall-prevention guidelines also identify gait and balance problems as key indicators for fall-risk assessment [2, 3]. Falls in older adults can lead to injury, fall-related morbidity, loss of independence, and substantial healthcare burden [1, 4]. These consequences underscore the need for accurate assessment and early identification of balance dysfunction.

Maintaining upright balance requires the coordinated integration of sensory input, central processing and motor output [5], relying on the coupling of the musculoskeletal system, neural conduction, and higher cognitive function. Aging is associated with declines in neuromuscular control, including changes in muscle strength, activation patterns, and co-contraction strategies [6–8]. During challenging balance tasks, older adults may show altered muscle activation and coordination strategies [7, 9], while age-related cognitive declines necessitate greater prefrontal involvement to maintain stability. Under increased cognitive load, limited cognitive resources may further compromise postural control, particularly in older adults with cognitive impairment [10]. These changes collectively suggest a neuro-muscular-cognitive profile associated with balance dysfunction, underscoring the need for multimodal assessment and integrative modeling.

In clinical and community contexts, balance and fall risk are commonly evaluated using established tools, such as the Berg Balance Scale (BBS), the Tinetti Performance-Oriented Mobility Assessment (POMA), and the Timed Up and Go (TUG) test [3, 11–13]. Although practical and widely used, these assessments are rater-dependent and have limited resolution. They fail to elucidate the neuro-muscular-cognitive mechanisms underlying balance control, including neuromuscular recruitment characteristics, muscle synergy organization, postural control dynamics, and cognitive load. Together, these objective physiological measures provide complementary information about dysfunction-related control patterns, with muscle synergies offering additional insight into how multiple muscles are organized as functional modules during postural control. Such information may help reveal functional heterogeneity that cannot be fully captured by clinical scores alone [7]. Individuals with similar clinical scores may therefore exhibit different underlying neuro-muscular-cognitive response patterns, masking functional heterogeneity that may be relevant for individualized interpretation.

Wearable sensors and machine learning offer a path toward more objective, scalable assessment. For example, random forest models with tri-axial accelerometry can predict fall risk [14], CNN-LSTM architectures with inertial units improve activity recognition accuracy [15] and one-dimensional convolutional neural network (1D-CNN) has also been used to analyze gait features [16]. Nevertheless, current approaches often rely on single-modality data, particularly kinematics, and lack the multimodal depth needed to capture neuro-muscular-cognitive coupling. They also often fail to account for population heterogeneity, resulting in limited ability to characterize heterogeneity in dysfunction-related functional patterns or to support more individualized interpretation of balance dysfunction.

To address these limitations, we leverage multimodal data to characterize the neuro-muscular-cognitive features associated with balance dysfunction in older adults. By integrating cortical activation, neuromuscular activity, postural control dynamics, and cognitive performance, this approach aims to provide a more objective and comprehensive description of balance-related functional changes under different postural demands. We propose an interpretable, two-stage supervised and unsupervised assessment framework based on multimodal data to classify dysfunction and explore balance-related heterogeneity.

Materials and methods

Participants

A total of 81 community-dwelling elderly individuals aged 60 years and over were recruited in this study. Following the BBS criterion that a score ≤ 45 indicates high fall risk and balance dysfunction [12], participants were divided into a balance dysfunction group (n = 38; BBS ≤ 45) and a healthy control group (n = 43; BBS > 45). All participants were right-leg dominant.

Inclusion criteria: (1) age ≥ 60; (2) stable vital signs, no severe visual and hearing impairment, able to understand and follow instructions to complete the assessment tasks; (3) no history of major surgery in the past 6 months, able to walk independently. Exclusion criteria: (1) severe bone and joint diseases, spinal lesions, arthritis, visual impairment, severe cardiovascular diseases; (2) related nervous system diseases, such as Parkinson’s disease, stroke, etc.; (3) potential sensory system and vestibular system diseases. This study was approved by the hospital ethics committee (No. SBKT-2024-028), and all subjects signed an informed consent form before participating.

Experimental design and procedure

Clinical scale assessment

Balance was assessed using the BBS and used for group allocation. Global cognition was evaluated with the Montreal Cognitive Assessment (MoCA), including domain scores for visuospatial, executive function, attention, memory, language, abstraction, calculation, and orientation [17].

Experimental postural tasks

Biomechanical assessments consisted of six static standing tasks, including four double-leg tasks and two single-leg tasks (Fig. 1). The double-leg tasks were natural stance (NS), feet-together stance (TS), right tandem stance (RTS), and left tandem stance (LTS). For NS, participants stood naturally with both feet placed comfortably apart. For TS, participants stood with both feet together. For tandem stance, one foot was positioned directly in front of the other, with either the right foot (RTS) or the left foot (LTS) leading. Each double-leg posture was maintained for 30 s with a 15 s rest interval between trials. Participants remained in a quiet relaxed standing position during the rest period, which served as the baseline epoch for the subsequent fNIRS task contrast. Although both the rest period and NS involved natural bipedal standing, NS was treated as a formal task epoch requiring stable standing during synchronous multimodal recording. The single-leg tasks included right and left single-leg stance (RSS and LSS), each with a target duration of 15 s. For RSS and LSS, participants were instructed to stand on the designated limb while maintaining balance as steadily as possible. If a participant could not maintain the posture for the full duration, the actual time sustained was recorded. Each task was performed twice under the same testing conditions. The mean value of the two trials for each task was used for subsequent analyses.

Fig. 1.

Experimental setup and balance tasks. a sEMG electrode placement; b Experiment environment; c Regions of interest; d Assessment tasks; e Task procedure. NS, natural stance; TS, feet-together stance; RTS, right tandem stance; LTS, left tandem stance; RSS, right single-leg stance; LSS, left single-leg stance

Data acquisition and processing

We performed synchronous multimodal acquisition during all balance tasks, recording center of pressure (COP), surface electromyography (sEMG), and functional near-infrared spectroscopy (fNIRS) signals. Preprocessing and feature extraction were implemented in MATLAB R2017b.

Functional near-infrared spectroscopy

Near-infrared functional brain imaging system (Nirsmart; Danyang Huichuang Medical Equipment Co., Ltd., Jiangsu, China) was used to monitor the changes of cortical blood oxygen in real time. It emitted near-infrared light with a sampling rate of 11 Hz. The probe was arranged according to the international 10–20 system with a fixed 3 cm source-detector separation, yielding 61 valid measurement channels. Based on task relevance, four bilateral regions of interest (ROI) were defined: the prefrontal cortex (LPFC/RPFC), supplementary motor area (LSMA/RSMA), premotor cortex (LPMC/RPMC), and primary motor cortex (LM1/RM1).

Preprocessing was performed with the Homer2 toolbox [18]. Raw light intensity was converted to optical density (OD), motion artifacts were identified and corrected using spline interpolation, and OD signals were band-pass filtered (0.01–0.10 Hz) to attenuate baseline drift and physiological noise. Using the modified Beer-Lambert law, filtered OD was transformed into relative concentration changes of oxygenated hemoglobin (HbO₂) and deoxygenated hemoglobin (HbR) [19]. In accordance with the block-design structure of the fNIRS protocol, the rest period was segmented as a separate baseline epoch for the subsequent task period. This task-rest contrast allowed task-evoked cortical oxygenation changes to be quantified while reducing the influence of inter-individual differences in resting cortical activity. Because the rest period and NS both involved natural bipedal standing, the NS-baseline contrast was interpreted cautiously as a relative change from relaxed baseline to formal task performance, rather than as activation induced by a distinct postural transition. For each ROI, the hemodynamic response function (HRF) was computed as the task-epoch average relative to the adjacent baseline.

Surface electromyography

A wireless sEMG system (MyoMove-COW, NCC Medical Co., Ltd., Shanghai, China) was used to record at 1000 Hz from eight bilateral lower-limb muscles: rectus femoris (RF), biceps femoris (BF), tibialis anterior (TA), and medial gastrocnemius (GM). The raw sEMG signal was filtered with a fourth-order Butterworth bandpass filter from 20 to 450 Hz, followed by full-wave rectification. A linear envelope was then obtained using a 50-sample moving Root mean square (RMS) window. For each trial, amplitude normalization was performed by dividing the linear envelope of each muscle by its peak value recorded within that trial, expressing muscle activation as a percentage. To enable temporal alignment across trials and participants, each trial was time-normalized and resampled to 101 equally spaced data points (representing 0–100% of the cycle) using linear interpolation, and subsequently smoothed using cubic B-splines.

Center of pressure

The COP data were acquired at 100 Hz using a force plate (Kistler, Winterthur, Switzerland). The raw COP time-series were filtered using a fourth-order zero-phase low-pass Butterworth filter with a cutoff frequency of 10 Hz to remove high-frequency noise without introducing phase distortion. To improve comparability across trials and participants, a consistent analysis window was extracted for each trial. In addition, the COP trajectories were mean-centered by subtracting the average mediolateral (M-L) and anteroposterior (A-P) COP positions from their respective time-series. The resulting M-L and A-P COP trajectories were then used to compute subsequent COP-related indicators.

Data analysis

fNIRS

Given that changes in oxygenated hemoglobin (ΔHbO₂) offer a higher signal-to-noise ratio than deoxygenated hemoglobin and are more sensitive to local perfusion and movement-related modulations [20], we used ΔHbO₂ as the primary HRF index. For each ROI, channel time series were averaged to obtain an ROI response. Baseline was defined as the resting epoch immediately preceding each task, and the mean task-epoch HRF amplitude relative to baseline was submitted to subsequent statistical analyses.

sEMG

To characterize muscle activation magnitude, antagonist-synergist relations, and postural control strategies, we computed the following indices:

(1) Root mean square (RMS): for each trial, RMS was computed from the normalized linear envelope x_i of the sEMG:

1

where x_i represents the normalized sEMG amplitude at the i-th sample point, and N denotes the total number of samples in the resampled time-normalized envelope.

(2) Contribution rate (Conrate): using the area under the linear envelope (AUC) as an amplitude surrogate, the relative contribution of muscle m was calculated as

2

where AUC_m represents the integrated area of muscle m, AUC_k is the integrated area of the k-th muscle, and M denotes the total number of muscles included in the calculation.

(3) Co-contraction index (CCI): following Falconer and Winter [21]

3

Where represents the overlapping activation area of the antagonist muscle pair, calculated as the integral of the lower value of the two normalized sEMG envelopes at each time point across the task duration. denotes the integral of the summed activation of the two muscles across the entire task.

In addition, non-negative matrix factorization (NMF) was used to extract muscle synergies across tasks [22, 23]. The pre-processed 8-channel sEMG data matrix (), consisting of the smoothed linear envelopes extracted from each muscle, was decomposed into a set of synergy modules, each containing a spatial component (W) and a temporal activation coefficient (H). The sEMG data matrix () is reconstructed by linear combination with the formula

4

Where e is the residual. In this study, the number of synergies was increased from 1, and the Variance account for (VAF) was used to test the reconstruction accuracy. The formula for the calculation of VAF is:

5

Where m is the number of samples. When VAF exceeds 90%, the reconstruction accuracy was considered to be high enough. The minimum number of synergies above 90% is determined, then the number of synergies is determined, and the muscle weight W and activation curve H are saved.

COP

Based on the COP trajectories in different task conditions, the following metrics were calculated to assess postural control ability:

(1) Oscillation range:

6

7

(2) Velocity:

8

9

(3) Area: The COP sway area was estimated using the polygon area method based on the sequence of COP trajectory points.

10

where x_i and y_i represent the mediolateral and anteroposterior COP coordinates at the i-th sampling point, N denotes the total number of sampling points, and T denotes the total duration of the trial. This metric reflects the overall spatial dispersion of the COP trajectory during the standing task.

Machine-learning-based classification and clustering

Building on a multimodal feature set, cortical activation, sEMG metrics, COP indices, and cognitive scores, this study adopted a two-stage, application-oriented analytical workflow. In the first stage, supervised classification was used to classify task-specific multimodal feature profiles according to balance dysfunction-related status. In the second stage, unsupervised clustering was applied to the profiles predicted as dysfunction in order to further explore within-group heterogeneity across postural task contexts.

Feature extraction

To construct the machine learning inputs, features showing significant between-group differences in the statistical analyses described in Sect.  2.4 were selected as candidate predictors. The selected features included multimodal indicators from four domains:

1. fNIRS: ROI-averaged ΔHbO₂ signals from cortical regions including the PMC, SMA, and PFC;

2. sEMG: RMS and conrate of the RF, BF, and GM, as well as selected CCI between antagonist muscle pairs and the number of extracted muscle synergy modules;

3. COP: A-P oscillation range, A-P and M-L velocity, and overall area;

4. Cognition: MoCA total score and selected subdomain scores (language, attention, calculation, and memory).

Candidate features were derived from all six postural conditions, and covered the four domains described above. In the final analysis framework, significance-based feature filtering was performed strictly within the training data to avoid information leakage, and the selected features were then used to train the classifiers. Across the implemented classification pipeline, 33 multimodal features constituted the final input set. All features were scaled to [0,1] using a MinMaxScaler to mitigate unit and magnitude differences.

Construction of the classification model

To predict the presence of balance dysfunction, three predefined classifiers were trained and compared on the same multimodal feature set: support vector machine (SVM), random forest (RF), and extreme gradient boosting (XGBoost). Data were split by stratified random sampling into 70% training and 30% held-out test sets. Hyperparameters were optimized using 5-fold cross-validated within the training set, and the final selected hyperparameter configuration for each classifier is summarized in Table 1. The held-out test set was used only for final performance reporting and was not used for hyperparameter tuning. Model performance was evaluated using accuracy and area under the receiver operating characteristic curve. The classifier with the strongest overall performance was used for subsequent model interpretation and for identifying dysfunction-related profiles for clustering. A supplementary 10-fold cross-validation analysis was conducted to further assess classifier stability. Feature importance was assessed using model-based importance scores, and SHAP (Shapley Additive Explanations) analysis was further applied to interpret the relative contribution of individual predictors to the final model output. A supplementary sensitivity analysis was also performed using a SHAP-guided reduced feature subset.

Table 1.

Final selected hyperparameter values for each classifier

Machine learning model	Hyperparameter	Final value
RF	Number of trees	100
	Splitting criterion	Entropy
	Maximum depth of tree	None
	Minimum number of samples to split an internal node	5
	Minimum number of samples at a leaf node	1
SVM	Kernel function	Linear
	Regularization penalty parameter	0.1
	Kernel coefficient	Scale
XGBoost	Number of base learners	100
	Learning rate	0.2
	Maximum depth of tree	5
	Subsample rate	1
	Feature sampling rate	0.8

Construction of the clustering model

Unsupervised K-means clustering was applied to task-specific multimodal feature profiles derived from participants predicted as balance dysfunction by the classification model, in order to explore within-group heterogeneity across postural task contexts. Prior to clustering, all features were standardized using z-score normalization (StandardScaler). The optimal number of clusters was evaluated using the elbow method based on the within-cluster sum of squares (WCSS), with candidate values ranging from k = 1 to k = 5, and the optimal solution was identified as k = 3. The final K-means model was implemented with k = 3, random state = 42, and 10 random initializations to reduce the risk of convergence to local minima. Cluster validity was further assessed using the silhouette coefficient. For visualization, normalized features were projected to two dimensions via principal component analysis (PCA). Cluster-wise feature means were then computed to compare profiles across clusters and characterize clustering-derived functional patterns. To further evaluate clustering robustness, a supplementary stability analysis was performed using repeated 80% subsampling. In each of 1000 iterations, K-means clustering with k = 3 was refitted to the subsampled data, and the agreement between the original clustering solution and the refitted labels for overlapping observations was quantified using the adjusted Rand index (ARI). Cluster-wise stability was additionally summarized using Jaccard similarity after optimal cluster-label matching.

Statistical analysis

SPSS 25.0 (IBM SPSS Statistics, Chicago, IL, USA) was used for statistical analysis. Normality was assessed with the Shapiro-Wilk test. Continuous variables are reported as mean ± SD. Group differences were examined using a two-way repeated-measures ANOVA with group as the between-subject factor and task as the within-subject factor. When the sphericity assumption was violated, the Greenhouse-Geisser correction was applied. Significant effects were followed by Bonferroni-adjusted post hoc comparisons. Effect sizes for ANOVA results were reported as partial eta squared (η_p²). For fNIRS analyses, repeated-measures ANOVA was conducted separately for each ROI. Bonferroni-adjusted post hoc comparisons were applied within each ROI, and the interaction-effect P values across ROIs were further controlled using false discovery rate (FDR) correction. Statistical significance was set at α = 0.05.

Results

Demographic characteristics

Table 2 summarizes demographic and scale-based characteristics. Compared with healthy older adults, the balance dysfunction group was older and showed significantly lower BBS and MoCA total scores (P < 0.001). Sex distribution, height, weight, and BMI did not differ significantly between groups.

Table 2.

Characteristics of the participants

Age, years	Balance dysfunction elderly (n = 38)	Healthy elderly (n = 43)	P value
	74.79 ± 8.25	69.77 ± 5.33	0.002*
Gender (male/female), n	11 / 27	15 / 28	0.574
Height, cm	159.00 ± 7.98	160.56 ± 7.72	0.375
Weight, kg	59.67 ± 7.85	62.47 ± 11.42	0.199
BMI, kg/m2	23.62 ± 2.81	24.14 ± 3.40	0.460
BBS	41.68 ± 3.89	52.56 ± 2.45	0.000***
MoCA	20.50 ± 7.54	25.30 ± 2.89	0.000***

Note. Values are mean ± SD, *: P < 0.05; ***: P < 0.001

COP parameters

Across tasks, A-P and M-L COP oscillation range, velocity, and area are summarized in Table S1. A significant group × task interaction emerged for A-P oscillation range (F = 6.244, P < 0.001, η_p² = 0.294). Post hoc tests showed that the balance dysfunction group exhibited greater A-P oscillation range than controls during TS, RSS, and LSS. Significant group × task interactions were also found for COP area (F = 4.088, P < 0.05, η_p² = 0.294) and velocity in both the A-P (F = 9.604, P < 0.001, η_p² = 0.403) and M-L (F = 6.101, P < 0.001, η_p² = 0.301) directions, with the dysfunction group showing higher values primarily under single-leg stance conditions. Specifically, larger area and higher M-L/A-P velocity were observed in RSS and LSS in the dysfunction group. Although no interaction was detected for the M-L oscillation range, both the task main effect and group main effect were significant, indicating that M-L oscillation range increased with task difficulty in both groups and remained overall greater in the dysfunction group. Overall, the dysfunction group demonstrated larger COP excursions across challenging postural tasks, whereas the healthy group showed comparatively smaller changes across task conditions.

Muscle activation parameters

RMS

Figure 2 illustrates RMS across muscles and tasks for both groups, and detailed ANOVA and post hoc results are provided in Table S2. For RF, significant group × task interactions were observed for both LRF (F = 2.954, P < 0.05, η_p² = 0.165) and RRF (F = 5.290, P < 0.001, η_p² = 0.261). Post hoc comparisons showed that the healthy group exhibited higher LRF activation during TS and higher RRF activation during RTS. During single-leg stance tasks, the dysfunction group showed relatively higher RF activation in the stance limb, reflected by greater RRF activation during RSS and greater LRF activation during LSS. Significant group × task interactions were also observed for BF (LBF: F = 2.439, P < 0.05, η_p² = 0.140; RBF: F = 3.848, P < 0.05, η_p² = 0.204). Specifically, the dysfunction group displayed greater BF RMS in multiple tasks (NS, TS, LTS, RSS on the right; NS, LSS on the left). Significant group × task interactions were additionally observed for TA (LTA: F = 5.316, P < 0.001, η_p² = 0.262; RTA: F = 4.285, P < 0.05, η_p² = 0.222), and post hoc comparisons indicated that the dysfunction group generally showed higher TA activation during single-leg stance tasks, whereas the healthy group exhibited comparatively more stable activation patterns across tasks.

Fig. 2.

RMS value of surface electromyography signal in different tasks. RF, rectus femoris; BF, biceps femoris; TA, tibialis anterior; GM, gastrocnemius; L, left; R, right; NS, natural stance; TS, feet-together stance; RTS, right tandem stance; LTS, left tandem stance; RSS, right single-leg stance; LSS, left single-leg stance

CCI

For CCI, a significant main effect of task was observed in all four muscle pairs, whereas neither the group main effect nor the group × task interaction reached significance (Table S2, P > 0.05). Although the trend plots (Fig. S1) suggested a steeper decline in thigh CCI with increasing task difficulty in the dysfunction group, these between-group differences did not reach statistical significance.

Muscle synergy

To explore differences in neuromuscular coordination organization, we compared the number of extracted muscle synergies across tasks (Table S3). In bipedal stances, the two groups did not differ significantly in synergy count. In single-leg stances, however, the balance dysfunction group exhibited a greater number of synergies than controls.

Figure 3 illustrates representative synergy weights (W) and activation coefficient curves (H). During bipedal stances, the healthy group’s Synergy 1 reflected a more balanced, multi-muscle participation, whereas the dysfunction group showed elevated weights for the LTA and LGM. In Synergy 2, the dysfunction group continued to show greater weighting of LTA and LGM, with lower weights for RF and BF. By contrast, the healthy group maintained a more even distribution of weights, with only modest increases in selected calf muscles. The dysfunction group also showed mild fluctuations in H between 30% and 50% of the task duration. In single-leg stances, compared with controls, the dysfunction group demonstrated higher contralateral weights (LBF and LGM) when standing on the right leg. During left single-leg stance, the last 20% of the task was characterized by higher LRF and LTA weights and lower BF and GM weights.

Fig. 3.

Muscle synergy patterns. RF, rectus femoris; BF, biceps femoris; TA, tibialis anterior; GM, gastrocnemius; L, left; R, right; NS, natural stance; TS, feet-together stance; RTS, right tandem stance; LTS, left tandem stance; RSS, right single-leg stance; LSS, left single-leg stance

Cortical activation

Figures 4 and S2 depict between-group differences and spatial maps of ROI-averaged ΔHbO₂, and detailed ANOVA and post hoc results are provided in Table S4. Significant group × task interactions were observed for LPMC (F = 4.011, P < 0.05, η_p² = 0.092), RPMC (F = 3.709, P < 0.05, η_p² = 0.083), LSMA (F = 5.967, P < 0.001, η_p² = 0.092), RSMA (F = 7.004, P < 0.001, η_p² = 0.112), LPFC (F = 3.936, P < 0.05, η_p² = 0.073), and RPFC (F = 3.357, P < 0.05, η_p² = 0.060), and all remained significant after FDR correction across ROIs. Overall, the dysfunction group exhibited higher cortical activation across multiple tasks, with the most pronounced effects in the PMC and SMA. Specifically, within PMC, LPMC during TS and RPMC during NS, TS, LTS, and RSS were significantly higher in the dysfunction group than in controls. Within SMA, both LSMA and RSMA were significantly elevated in the dysfunction group during NS, TS, and LTS. In M1, LM1 showed a significant group main effect (F = 4.686, P < 0.05, η_p² = 0.016), with higher activation in the dysfunction group during TS, RSS, and LSS, whereas RM1 showed no significant group difference. Finally, in the PFC, LPFC was higher in the dysfunction group across all tasks, whereas RPFC was higher during NS, TS, LTS, and RSS.

Fig. 4.

Group differences of ROI-averaged ΔHbO₂. NS, natural stance; TS, feet-together stance; RTS, right tandem stance; LTS, left tandem stance; RSS, right single-leg stance; LSS, left single-leg stance; PMC, premotor cortex; SMA, supplementary motor area; PFC, prefrontal cortex; M1, primary motor cortex; L, left; R, right

Machine-learning-based classification and clustering

Classification

Figure 5 presents confusion matrices and ROC curves for the three classifiers. Overall performance is summarized in Table 3. XGBoost showed the strongest overall performance, with 83.56% held-out test accuracy, exceeding RF (74.66%) and SVM (70.55%), and its AUC (0.87) also surpassed RF (0.83) and SVM (0.81).

Fig. 5.

Confusion matrices and ROC curves. a Support vector machine (SVM); b Random forest (RF); c XGBoost. Dys, balance dysfunction group; Hea, Health control group

Table 3.

Performance comparison across classifiers

Model	Accuracy (%)	AUC
SVM	70.55	0.81
RF	74.66	0.83
XGBoost	83.56	0.87

Feature attribution for XGBoost (Fig. 6 and S3) indicated high contributions from COP metrics, particularly velocity in both the A-P and M-L directions as well as COP oscillation range. Notably, M-L velocity ranked among the most influential COP features in the SHAP analysis. Additional contributions were observed from cognitive domains (attention, memory, calculation), sEMG (RBF, RRF, LRF, RGM), and fNIRS (LSMA, LPMC, RPFC). SHAP visualizations showed that higher RF activation and higher MoCA totals and sub-domain scores shifted predictions toward healthy status, whereas elevated RGM and LBF activation and greater LSMA, LPMC, and RPFC activation shifted predictions toward balance dysfunction. A supplementary sensitivity analysis using a SHAP-guided reduced feature subset revealed model-specific differences in response to feature compression (Table S5). Relative to the original feature set, accuracy increased from 70.55% to 82.19% for SVM and from 74.66% to 77.40% for RF, whereas XGBoost decreased from 83.56% to 78.77%, while still maintaining competitive performance. A supplementary 10-fold cross-validation analysis showed that XGBoost retained the best overall stability (mean accuracy: 0.8353 ± 0.0606; mean AUC: 0.9117 ± 0.0484), followed by RF (0.8176 ± 0.0718; 0.8897 ± 0.0607) and SVM (0.7588 ± 0.0600; 0.8100 ± 0.0644), further supporting the robustness of the classification results.

Fig. 6.

Feature importance. a Muscle activation; b Cortical activation; c COP indexes; d Cognitive domain. RF, rectus femoris; BF, biceps femoris; TA, tibialis anterior; GM, gastrocnemius; PMC, premotor cortex; SMA, supplementary motor area; PFC, prefrontal cortex; M1, primary motor cortex; L, left; R, right

Clustering-based pattern identification

Clustering analysis was then performed on task-specific multimodal feature profiles derived from participants predicted as dysfunction by the classification model to further examine within-group heterogeneity. The clustering dataset comprised 222 task-specific multimodal feature profiles derived from predicted-dysfunction participants, and individual participants could contribute feature profiles to different clusters across task conditions. K-means clustering indicated an optimal cluster number of k = 3 according to the elbow criterion. The final clustering solution contained 134, 49, and 39 feature profiles in Cluster 0, Cluster 1, and Cluster 2, respectively. In the PCA projection (Fig. S4), the three clusters showed partial visual separation, with a silhouette coefficient of 0.1553 (Table S6). The supplementary cluster stability analysis using repeated 80% subsampling showed an average ARI of 0.792 ± 0.108. The mean cluster-wise Jaccard similarity was 0.849 ± 0.126, with values of 0.891 ± 0.062, 0.922 ± 0.062, and 0.732 ± 0.140 for the three clusters, respectively (Table S7).

Cluster-wise means showed significant differences in BBS and MoCA scores across clusters (P < 0.001). Based on their relative functional and cognitive characteristics, the three clusters were descriptively labeled as higher-decline, intermediate, and relatively preserved patterns (Table 4). The high-decline pattern exhibited higher bilateral BF and GM activation, whereas the relatively preserved pattern showed higher bilateral RF activation. In fNIRS, the high-decline pattern demonstrated elevated activation in LPFC/RPFC, LPMC/RPMC, and LSMA/RSMA. In cognition, the relatively preserved pattern showed the best overall MoCA performance (Fig. 7).

Table 4.

Key feature means by cluster

BBS	High-decline group	Intermediate group	Relatively preserved group
	36.51	38.84	41.16
MoCA	11.31	19.96	22.96

Fig. 7.

Cluster-wise feature comparisons. RF, rectus femoris; BF, biceps femoris; TA, tibialis anterior; GM, gastrocnemius; PMC, premotor cortex; SMA, supplementary motor area; PFC, prefrontal cortex; M1, primary motor cortex; L, left; R, right

Discussion

By integrating fNIRS, sEMG, COP, and cognitive scores, this study identified a multimodal feature pattern associated with balance dysfunction in older adults. Under more challenging postural conditions, the dysfunction group showed impaired postural control (increased COP metrics), accompanied by greater muscle activation, altered synergy organization, and increased cortical activation in prefrontal and motor-related regions. Together with concurrent cognitive decline, these findings suggest that balance dysfunction in older adults is not merely a peripheral motor problem, but involves broad alterations across motor, cortical, and cognitive domains, some of which may reflect adaptive responses to increased postural demand. Building on these findings, we developed a multimodal classification-and-clustering framework, in which XGBoost achieved a classification accuracy of 83.56%, and K-means clustering further identified three functionally distinct patterns with graded differences, which differed in BBS, MoCA, and neuro-muscular-cortical characteristics. Overall, this framework not only helps characterize balance dysfunction using objective physiological signals, but also provides a new perspective for understanding its internal heterogeneity.

The dysfunction group showed significantly lower scores on both BBS and MoCA, with notable deficits in attention, memory, and language. These findings suggest that balance dysfunction in older adults is not solely a peripheral motor problem, but may also involve reduced cognitive resources for postural monitoring and control [24]. Within this motor-cognitive framework, limited attentional capacity and impaired executive function may compromise the real-time allocation of resources required for maintaining stability, particularly under more demanding postural conditions [25]. This interpretation is consistent with the broader multimodal pattern observed in the present study, in which cognitive decline co-occurred with increased cortical activation, altered neuromuscular recruitment, and impaired postural control. Together, these findings support the inclusion of cognitive assessment in balance evaluation and highlight the relevance of motor-cognitive interactions in balance dysfunction.

Across tasks, the dysfunction group demonstrated impaired postural stability, evidenced by increased oscillation range, velocity, and area, particularly during challenging conditions like single-leg stance, indicating pronounced deficits in A-P control [7, 26, 27]. This A-P instability may stem from reduced knee extensor strength and impaired anticipatory control [29, 30]. In addition, although M-L oscillation range did not show a significant interaction effect, M-L velocity was significantly elevated in the dysfunction group under more demanding postural conditions, suggesting that mediolateral control deficits may be reflected more in dynamic sway regulation than in sway amplitude itself. Since single metrics such as stance time may overlook subtle deficits [28], graded task conditions combined with multimodal measures may help detect more nuanced dysfunction-related patterns.

We observed reduced recruitment efficiency of antigravity musculature in the dysfunction group, with greater muscle activation under high-demand tasks. Specifically, RF activation was lower than that of the healthy group in tasks requiring knee extensor support, whereas BF and GM showed higher activation in several more challenging tasks. This pattern is broadly consistent with Nagai et al.’s report of increased ankle-strategy involvement in older adults [29]. Rather than indicating a generalized increase in co-contraction, these findings may reflect a shift toward more localized or task-specific recruitment strategies [30]. In addition, during tasks such as RTS and LSS, the dysfunction group showed greater fluctuations in TA and GM activity around the ankle, which may indicate less stable neuromuscular control, although such variability may also be related to the increased postural demands of maintaining unstable positions [9].

Contrary to previous studies reporting greater co-contraction in older adults during postural control tasks [9, 31], we did not observe a significant increase in CCI in the dysfunction group. This inconsistency may reflect differences in comparison groups, task demands, and muscle pairs analyzed. Whereas previous studies often compared healthy older adults with young adults, the present study focused on balance-dysfunction older adults versus healthy older adults, and may therefore capture within-age-group heterogeneity rather than aging effects alone. Co-activation may also vary across body regions and task contexts rather than increase uniformly across all muscles. Accordingly, our findings suggest that neuromuscular adaptation in this cohort may be expressed more through selective muscle overactivation than through a generalized increase in co-contraction.

Consistent with this interpretation, the dysfunction group showed a greater number of extracted muscle synergies under high-difficulty tasks and greater reliance on local muscle groups, suggesting a less flexible coordination pattern relative to that observed in healthy peers [7, 32]. In the context of the present static postural tasks, these synergy-related differences should be interpreted cautiously as exploratory representations of altered neuromuscular organization, rather than direct equivalents of the phasic synergy structures commonly reported in dynamic tasks such as gait. Moreover, the observed changes in synergy weights and activation coefficients indicate that the dysfunction group may require more task-specific redistribution of muscle recruitment to maintain posture under demanding conditions. Additionally, greater within-task variability of activation time series may reflect reduced stability and precision in neuromuscular regulation, although it may also be related to the increased postural demands required to maintain a more unstable posture.

Across multiple tasks, the dysfunction group showed greater activation in PMC, SMA, and PFC, particularly under high task difficulty. This pattern may reflect increased neural recruitment associated with maintaining postural control under greater demand [33]. From a neurophysiological perspective, PMC and SMA are involved in motor planning and postural coordination across multiple joints. As task demands increase, greater engagement of premotor and supplementary motor regions may support the coordination of muscle activation sequences and inter-joint regulation [34]. The increased activation of these regions in older adults with balance dysfunction relative to healthy older adults may reflect less efficiency and less local control strategies. In addition, increased prefrontal activation may indicate greater involvement of attentional and executive resources during balance control. Under more complex conditions, however, this increased cortical recruitment may not be sufficient to fully stabilize posture, as suggested by the co-occurrence of increased COP instability and altered muscle synergy patterns [35].

In the classification task, XGBoost demonstrated the strongest overall performance, consistent with its ability to capture higher-order feature interactions [36]. Notably, although SVM achieved an acceptable AUC, indicating reasonable probabilistic ranking, its limited class separation at the default decision threshold resulted in a nearly balanced confusion matrix (Fig. 5a). Tree-based models such as RF and XGBoost therefore appear better suited to capturing the complex nonlinear relationships among the multimodal features considered here. The supplementary 10-fold cross-validation analysis suggested that the relative performance of XGBoost was reasonably stable across data partitions, although external validation remains necessary. This may reflect the ability of tree-based ensemble models to capture nonlinear interactions among multimodal features [36].

Feature-importance and SHAP analyses clarified the physiological drivers of model decisions. Among COP variables, velocity in both the anterior-posterior and mediolateral directions, together with oscillation range, emerged as important discriminators. Notably, M-L velocity was not only highly ranked in the machine-learning model, but also showed significant task-dependent group differences in the conventional biomechanical analysis, particularly under single-leg stance conditions. This finding suggests that mediolateral sway control may play an important role in distinguishing balance dysfunction, especially when considered jointly with neuromuscular and cognitive features. In the cognitive domain, memory, calculation, and attention carried prominent weights, underscoring how cognitive decline can exacerbate postural control deficits. For sEMG features, higher contributions from BF and GM, with RF activation associated with the healthy class, are broadly consistent with a pattern of reduced extensor contribution alongside greater flexor/plantarflexor involvement. Finally, elevated importance of PMC, SMA, and PFC signals is concordant with the group-level finding of increased cortical recruitment in dysfunction. Collectively, these results enhance model interpretability and are consistent with the biomechanical and functional findings described above. Notably, the SHAP-guided reduced feature set achieved highly competitive classification performance compared to the original input set, while revealing differing model sensitivities to feature compression. Dimensionality reduction improved SVM’s accuracy but led to a slight decline in XGBoost’s performance. This demonstrates that, unlike SVM, which thrives in a streamlined feature space, XGBoost can effectively represent complex, high-order interactions within the data, further validating its distinct superiority and application potential in parsing multimodal features [36].

We further applied clustering analysis to task-specific multimodal feature profiles derived from participants predicted as balance dysfunction. This clustering procedure identified three exploratory functional patterns with graded differences in BBS and MoCA scores, suggesting potential heterogeneity within dysfunction-related multimodal profiles across postural task contexts. The low silhouette coefficient indicated limited geometric separation and partial overlap among clusters, suggesting that these patterns should not be interpreted as clearly discrete categories. Rather, they may represent partially overlapping response patterns along a continuum of balance dysfunction, consistent with the continuous nature and high-dimensional complexity of neuro-muscular-cognitive alterations in older adults. However, the supplementary subsampling-based stability analysis showed relatively high agreement between the original and refitted clustering solutions, indicating that the clustering structure was not entirely driven by a single data partition. Therefore, these clusters should be interpreted as exploratory but reasonably stable functional response patterns, rather than definitive clinical subtypes.

Clustering-derived pattern profiling revealed graded differences across the three exploratory functional patterns. Compared with the relatively preserved pattern, the higher-decline pattern showed greater muscle activation, poorer postural stability, increased cortical activation, and lower cognitive performance, whereas the intermediate pattern showed values between these two patterns. These findings suggest that the identified patterns may differ in the relative involvement of motor, cortical, and cognitive processes [37]. However, given the limited geometric separation among clusters, these patterns should be interpreted as partially overlapping functional response patterns across postural contexts, rather than discrete clinical subtypes. They may be useful for future hypothesis generation and intervention-oriented research.

In summary, our multimodal framework provides an interpretable approach for classifying balance dysfunction and exploring heterogeneity in dysfunction-related functional patterns, while offering richer physiological characterization than conventional scales alone. However, several limitations should be acknowledged. First, the framework relies on relatively complex and resource-intensive measurements, limiting its immediate applicability for community screening. Second, short-separation channels were not available in the fNIRS system used in this study. Therefore, superficial physiological signals could not be explicitly regressed out. This may limit the interpretation of cortical oxygenation changes, and the fNIRS findings should be interpreted with appropriate caution. Third, although the supplementary stability analysis suggested relatively high clustering consistency, the low silhouette coefficient indicated limited geometric separation among clusters. Therefore, the clustering-derived patterns should be regarded as exploratory and require validation in larger independent cohorts. Future studies should prioritize longitudinal validation, larger independent cohorts or nested cross-validation, comparison with simpler clinical tools, fNIRS correction strategies for superficial physiology, and the development of reduced sensor configurations to improve feasibility in real-world settings.

Conclusion

Based on multimodal physiological signals, this study identified a dysfunction-related neuro-muscular-cognitive profile in older adults. Under more challenging postural conditions, this profile was associated with reduced knee extensor contribution, greater activation of knee flexors and ankle plantarflexors, increased recruitment of PMC, SMA, and PFC, impaired postural control, and lower cognitive performance. These findings suggest that balance dysfunction in older adults involves coordinated alterations across motor, cortical, and cognitive systems. The proposed multimodal classification-and-clustering framework, integrating objective physiological features across cortical, neuromuscular, postural, and cognitive domains, provides an interpretable approach for dysfunction-related classification and for exploring heterogeneity in functional patterns beyond conventional clinical scales. This framework may also offer useful directions for future intervention-oriented research.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviations

A-P

Anterior-posterior

BBS

Berg balance scale

BF

Biceps femoris

CCI

Co-contraction index

COP

Center of pressure

fNIRS

Functional near-infrared spectroscopy

GM

Gastrocnemius

HbO₂

Oxygenated hemoglobin

LTS

Left tandem stance

LSS

Left single-leg stance

M-L

Medio-lateral

MoCA

Montreal cognitive assessment

M1

Primary motor cortex

NMF

Non-negative matrix factorization

NS

Natural stance

PFC

Prefrontal cortex

PMC

Premotor cortex

RF

Rectus femoris

RMS

Root mean square

ROI

Region of interest

RTS

Right tandem stance

RSS

Right single-leg stance

sEMG

Surface electromyography

SMA

Supplementary motor area

TA

Tibialis anterior

TS

Feet-together stance

Author contributions

J.S.: Conceptualization, Methodology, Software, Formal analysis, Investigation, Data curation, Visualization, Writing - original draft; C.L.: Investigation, Data curation, Data preprocessing, Review and editing; M.S.: Data curation, Review and editing; Y.G.: Experiment organization; Z.W.: Experiment organization; F.W.: Experiment organization; Y.L: Review and editing; W.N.: Conceptualization, Writing - review and editing, Project administration, Funding acquisition, Supervision.

Funding

This study was supported by National Key Research and Development Program of China (2023YEC3603702) and the Shanghai Innovative Medical Device Application Demonstration Project (23SHS05400-06).

Data availability

The data that support the findings of this study are available upon reasonable request from the authors.

Declarations

Ethics approval and consent to participate

The study was conducted in accordance with the Declaration of Helsinki and was approved by the ethics committee of Shanghai YangZhi Rehabilitation Hospital (SBKT-2024-028). All participants provided written informed consent prior to participation.

Consent for publication

Consent for publication were given by all participants.

Competing interests

The authors declare no competing interests.