Digital Balance Biomarkers and Interpretable Evaluation of Disease Severity in Spinocerebellar Ataxia Type 3

ABSTRACT

The clinical assessment of spinocerebellar ataxia type 3 (SCA3) is hindered by subjective factors and inter‐rater variability. This study utilizes a previously validated wearable plantar pressure insole system established in our prior work to build a multiscenario digital balance detection framework, and assesses its efficacy in differentiating SCA3 patients from healthy individuals and in evaluating disease severity. The study involved 74 SCA3 patients and 45 healthy controls, who were equipped with custom‐developed plantar force sensors to perform standing tasks with eyes open (EO) and eyes closed (EC). Center of pressure (COP) variables were analyzed to determine their correlation with clinical characteristics and their discriminative capability. An XGBoost model was employed to classify disease severity, with SHAP analysis providing enhanced interpretability. COP variables derived from EC tasks demonstrated superior performance in terms of area under the curve, accuracy, sensitivity, and specificity compared to those from EO tasks. Significant correlations were identified between COP variables, such as the eyes closed, left side sway velocity (EC‐L velocity) SD, and clinical scores (r > 0.4). The model achieved an accuracy of 87% in classifying disease severity, with SHAP analysis highlighting key variables and interactions. COP variables offer a robust, multidimensional measure of motor dysfunction.

Clinical assessment of spinocerebellar ataxia type 3 (SCA3) is hindered by subjectivity. This study uses a wearable plantar pressure insole to analyze center of pressure (COP) variables in 74 patients and 45 controls during standing tasks. Eyes‐closed COP indices yield better discriminative performance and correlate with clinical scores. The interpretable XGBoost model achieves 87% severity classification accuracy for objective SCA3 evaluation.

1. Introduction

Spinocerebellar ataxias (SCAs) are autosomal dominant neurodegenerative disorders characterized by genetic heterogeneity [1, 2]. To date, they have been linked to various causative genes and classified into 48 distinct subtypes. Spinocerebellar ataxia type 3 (SCA3), one of the most prevalent subtypes [3, 4], is characterized by a trinucleotide repeat expansion (cytosine‐adenine‐guanine, CAG) within the coding region of the gene ATXN3 located at chromosome 14q32.12, resulting in disruption of cellular pathways [5, 6]. The clinical manifestations of SCA3 exhibit significant heterogeneity. The principal pathological alterations are characterized by the progressive degeneration of the spinal cord and cerebellum, alongside dysfunctions in multiple systems, notably including lower motor neuron impairment. These pathological alterations intensify balance deficits, leading to gait abnormalities, compromised postural control, and movement incoordination [7, 8]. In advanced stages, patients may even lose the ability to stand independently [9]. Given the heterogeneous clinical manifestations of SCA3, objective and quantitative assessment of disease progression remains a major clinical challenge.

Current clinical workflows for SCA3 rely on genetic testing for definitive diagnosis, combined with the International Cooperative Ataxia Rating Scale (ICARS) and/or the Scale for the Assessment and Rating of Ataxia (SARA) to evaluate disease severity [10, 11, 12, 13]. However, these clinician‐rated scales are inherently limited by subjective judgment and inter‐rater variability, and cannot precisely quantify subtle motor and balance dysfunction. Smart devices such as optical tracking systems, force plate platforms, and full‐motion kinematic suits are constrained by high costs and time‐consuming procedures, limiting their application. In contrast, wearable pressure sensors are a superior option due to their portability and ability to provide diverse metrics [14, 15]. Plantar pressure–based temporal features have also been demonstrated to sensitively capture gait disturbances and treatment responses in neurological disorders; for instance, in idiopathic normal pressure hydrocephalus, a plantar‐pressure variation index derived from temporal alignment discriminated impaired from improved gait before and after cerebrospinal fluid tap test and shunt surgery [16]. Plantar pressure monitoring systems are categorized into platform‐based and in‐shoe systems [17]. The former, due to its bulky equipment and high cost, is primarily used in laboratory research [18]. While the latter facilitates capturing pressure distribution for stability analysis, its wired connection design compromises comfort and practicality. Recent evidence further indicates that plantar‐pressure distribution itself carries discriminative information for identifying ataxia, supporting the feasibility of pressure‐based objective assessment in neurological disorders [19].

Existing wearable‐based studies on SCAs mainly fall into two categories: exploring gait or balance features to distinguish ataxia patients from healthy populations [20], or applying machine learning to differentiate different SCA subtypes and other cerebellar degenerative disorders [21]. Previous posturographic studies have demonstrated that center of pressure (COP)‐derived measures, including sway velocity, sway area, and direction‐specific displacement, can discriminate SCA3 patients from healthy controls (HCs) and reflect postural instability [20]. Our prior study in the Journal of NeuroEngineering and Rehabilitation (JNER) demonstrated that combined wearable gait and static postural features effectively differentiated clinically overlapping SCA3 and cerebellar multiple system atrophy (MSA‐C) [22], validating the diagnostic value of gait plus COP biomarkers for cross‐subtype differentiation and identifying cadence variability and eyes‐closed (EC) COP metrics as core discriminative features. Despite the progress in wearable‐based differential diagnosis for ataxia, critical research gaps remain. First, most studies prioritize cross‐disease or cross‐subtype classification, with few systematically exploring static COP biomarkers for objective stratification of SCA3 severity (mild vs. moderate). Second, while COP parameters show group‐level differences between SCA3 patients and healthy individuals, their quantitative correlations with clinical rating scales, disease duration, age at onset, and other key clinical phenotypes remain incompletely characterized. Third, conventional statistical analyses only evaluate individual feature effects, neglecting interactive relationships between COP metrics and their combined impact on severity assessment. Additionally, machine learning models used in prior SCA3 research often lack interpretability, hindering clinical translation for routine severity monitoring.

To fill these gaps, the present study builds upon the wearable insole platform and well‐established participant cohort from our prior research [22]. Notably, the current paper differs in objective and analytical focus from the prior JNER work: the JNER study focused on multiclass differential diagnosis across SCA3, MSA‐C, and HCs using both gait and postural features, while the current work narrows the research scope to SCA3 alone, excludes gait‐related tasks and features, and exclusively focuses on static standing COP features and conduct a targeted analysis with three main objectives: (1) verify the discriminative efficacy of COP variables to distinguish SCA3 patients from age‐matched HC; (2) quantify the correlations between COP biomarkers and standard clinical ataxia scales and key clinical characteristics of SCA3; (3) construct an interpretable XGBoost model combined with SHAP analysis to stratify mild and moderate SCA3 patients, and reveal the independent contributions and interactive effects of core COP features. This work aims to provide objective, quantifiable digital balance biomarkers and a user‐friendly interpretable assessment framework to complement subjective clinical scales, supporting precise severity evaluation and individualized rehabilitation management for SCA3 patients.

2. Methods

2.1. Study Design and Participants

This research was conducted in compliance with the principles outlined in the Declaration of Helsinki. The study protocol, consent form, and associated documents received approval from the local ethics committee of The First Affiliated Hospital of Fujian Medical University, Fuzhou, China (MRCTA, ECFAH of FMU [2022]399), prior to the initiation of the study. Furthermore, the study conformed to the Consolidated Standards of Reporting Trials (CONSORT) guidelines. An independent data safety monitoring board evaluated the safety and quality of the study. SCA3 patients were recruited from the Organization in South‐East China for Cerebellar Ataxia Research (OSCCAR) in the Department of Neurology of the First Affiliated Hospital of Fujian Medical University (ClinicalTrials.gov: NCT04010214).

Informed written consent was obtained from all participants at the time of enrollment. The primary cohort for this study was recruited through hospital appointments or as relatives of patients. A total of 74 individuals with a molecular diagnosis of SCA3, along with 45 healthy participants, were enrolled in the study. Enrollment of patients was contingent upon the following inclusion criteria: (1) a confirmed genetic diagnosis of SCA3; (2) manifestation of ataxia symptoms; (3) voluntary consent to participate in the cohort study; (4) age of 18 years or older; and (5) a shoe size ranging from 36.5 to 44. The exclusion criteria employed for the selection of SCA3 patients were as follows: (1) the presence of known recessive, X‐linked and mitochondrial ataxias; (2) prior genetic tests excluding an SCA3 diagnosis; (3) asymptomatic patients and SCA3 homozygotes; (4) concurrent ailments that could impact SARA scores or be associated with balance impairments; (5) the inability to maintain an upright posture and normal gait without assistance; and (6) a demonstrated inability to comprehend and follow instructions. Meanwhile, HC individuals were selected to align with the patients in terms of age, gender, height, weight, and environmental characteristics. Individuals who were at risk of developing SCA3 due to their familial connection were excluded from the control group.

2.2. Objective Data Acquisition System

Figure 1A depicts the schematic diagram of the balance data acquisition and analysis flow in this study based on an insole‐based system. The system consists of a set of smart insoles integrated with a computer‐designed program. Each smart insole is equipped with a flexible plantar force sensor array that includes 20 sensing units (Figure S1) and a microcontroller. Figure 1B illustrates the structural composition of each sensing unit, which consists, from the bottom to the top, of a basal layer, an electrode layer, an air gap layer, a force‐sensitive layer, and an encapsulation layer. The 20 resistive force sensing units are strategically distributed across various regions of the foot (see Note S1). The force sensing units connect with an STM32F103C8T6 microcontroller for data acquisition at a sampling frequency of 20 Hz, preprocessing, and Bluetooth transmission to a computer for future feature extraction. The detailed preparation process, operational principles, and sensing performance of the smart insole have been thoroughly documented in a previous study [23].

FIGURE 1.

Schematic diagram of data collection and processing based on insole‐based systems: (A) balance data acquisition and analysis flow. (B) Structural composition of each force‐sensing unit in the smart insoles. (C) Two objective assessment tasks—natural standing with eyes open (EO, or Task 1) and eyes closed (EC, or Task 2). (D) Center of pressure (COP) and related variables extraction process. (E) Typical COP position change rate for normal subjects and patients during a 25‐s EC standing period.

2.3. Clinical and Objective Assessment

Genotyping was initially conducted to confirm the diagnosis of SCA3 and to ascertain the length of the CAG repeat expansion. Genomic DNA was extracted from peripheral blood samples using a standardized protocol with the QIAamp DNA Blood Minikit (QIAGEN, Hilden, Germany). Subsequently, polymerase chain reaction amplification was executed, and Sanger sequencing was used to determine the length of the CAG repeat sequence [24]. Following the completion of the genetic diagnosis, participants underwent concurrent clinical evaluations conducted by two experienced neurologists using both ICARS and SARA scales. Age at onset (AAO) refers to the age at which symptoms of SCA3 initially manifest. Disease duration is calculated as the interval between AAO and the age at which the patient first seeks medical consultation. The rate of disease progression is determined by dividing the SARA score by the disease duration.

Following the clinical assessment, a rehabilitation specialist (author X.‐Y.C.) blinded to the patients’ clinical phenotypes, performed an evaluation of the patients’ static balance on the same day. The participants wore the smart insoles, and the data acquisition and processing circuitry was securely affixed to the lower leg using nylon hook‐and‐loop straps. During the assessment, the participants were instructed to stand with their feet positioned naturally apart and their arms resting naturally at their sides. They were then required to perform two tasks: first, to maintain this posture with their eye open (EO, Task 1), and subsequently, to do so with their eye closed (EC, Task 2), each for a duration of 3 min as illustrated schematically in Figure 1C.

2.4. COP Variable Extraction

The procedure for extracting COP variables is illustrated in Figure 1D. Initially, during the assessment, plantar force data were transmitted in real‐time to the corresponding computational program via the smart insoles system. This system performed trimming and preprocessing to ensure uniformity in data shape and consistency in temporal spans. Following this, the preprocessed data, along with sensor positions, were synchronously input into the COP calculation model that predicted the participant's real‐time COP position and calculated associated variables (Figure 1E). Detailed COP calculation flow is presented in Note S2. Then, COP sway area and other dynamic characteristics, including the total sway area (A) of the COP, overall velocity (v), overall acceleration (a), and the velocity and acceleration of each foot in the medio‐lateral (ML) and antero‐posterior (AP) directions during the performance of standing tasks with EO and EC, were analyzed. Detailed definitions of these variables and their corresponding calculation equations are presented in Table S1.

2.5. Statistical and SHAP‐Based SCA3 Discrimination

All statistical analyses were performed with SPSS version 27.0. Initially, the normality of the quantitative variables was evaluated using the Kolmogorov−Smirnov test. Subsequently, the Kruskal−Wallis test was employed to compare differences between the SCA3 patient group and the HC group. In instances where the test results indicated a significant effect (p < 0.05), post hoc analysis was conducted using the Mann‒Whitney U test. To control for multiple comparisons among the 52 COP parameters in group difference analysis, the Bonferroni correction was applied. The area under the curve (AUC) was calculated to evaluate the discriminative efficacy of each parameter in distinguishing SCA3 patients from HC [25]. The Pearson correlation coefficient was employed to examine the relationship between clinical scales and balance parameters to discern subtle variations in balance ability. To account for multiple comparisons in the correlation analyses, p‐values were adjusted using the false discovery rate (FDR) method. Then, the Kruskal‒Wallis test and Mann−Whitney U test were utilized to analyze the correlation between the SARA scale and the subscores of the ICARS scale, respectively. In addition, Pearson's correlation coefficient was used to analyze the correlation between AAO, disease duration, disease progression rate, and high‐resolution parameters. The magnitude of group differences was quantified using Cohen's d [26]. A p‐value of < 0.05 was considered statistically significant, and a p‐value of < 0.01 was considered highly statistically significant [27].

To further assess the clinical status of patients with SCA3 and ascertain the disease's severity, we employed machine learning techniques, specifically extreme gradient boosting (XGBoost), in conjunction with SHAP values, to conduct an interpretable analysis of SCA3 severity based on the COP data. Detailed XGBoost model construction and evaluation are presented in Note S4.

2.6. Overlap With the Prior JNER Study

This section clarifies the overlap of participants, experimental protocols, data processing, and features between the current study and our prior research published in jner [22].

2.6.1. Participant Cohort

This study adopted the identical core cohort as the JNER work, including 74 genetically confirmed SCA3 patients and 45 age‐matched HCs. The JNER study additionally recruited 43 patients with MSA‐C for cross‐subtype differential diagnosis, and this MSA‐C cohort was not included in the present analysis.

2.6.2. Experimental Tasks

The EO and EC static standing trials (3 min per trial) were fully reused from the JNER study, including standing posture, participant instructions, and test environment. The 15‐meter straight walking gait trial implemented in the JNER study was excluded in this work, and no gait data were collected.

2.6.3. Data Acquisition and Preprocessing

The wearable insole hardware, sensor calibration protocol, raw plantar pressure data denoising, time synchronization, data trimming pipelines, and the COP position calculation algorithm were completely consistent with the prior JNER study, with no modifications applied.

2.6.4. Extracted Features

All 52 COP‐derived balance variables analyzed in this study belong to the COP feature subset of the 98 total features in the JNER study (full overlap for COP metrics). This study did not extract or utilize any gait‐related spatiotemporal features from the JNER dataset.

3. Results

3.1. Clinical Evaluation

The main cohort of this study consisted of 119 individuals, including 74 SCA3 individuals (49 males and 25 females) and 45 age‐matched HCs (27 males and 18 females). Table 1 presents the demographic characteristics of the study participants. The mean age of the SCA3 patients was 43.97 ± 10.54 years, with an average height of 164.06 ± 7.77 cm, a mean weight of 59.34 ± 10.88 kg, and a mean body mass index (BMI) of 21.86 ± 3.04. Statistical analysis revealed no significant differences between the patient and the HC group concerning gender (p = 0.495), age (p = 0.852), height (p = 0.251), weight (p = 0.175), or BMI (p = 0.164) distributions. The mean AAO for SCA3 patients was 37.11 ± 9.98 years. The mean disease duration was 7.38 ± 4.38 years, and the disease progression rate was 1.59 ± 1.05 SARA points per year. The median length of the CAG repeat sequence in the normal allele was 20.24 ± 6.96. Furthermore, SCA3 patients had mean scores of 8.96 ± 2.49 for SARA, 24.11 ± 7.18 for total ICARS, 2.07 ± 0.62 for the SARA eyes‐open subscore (SARA_EO), 4.6 ± 1.72 for the ICARS eyes‐open subscore (ICARS_EO), and 2.5 ± 1.12 for the ICARS eyes‐closed subscore (ICARS_EC). Based on the SARA score range, 17 patients were identified with mild ataxia, while 57 patients had moderate ataxia, with no instances of ataxia of other severities observed.

TABLE 1.

Demographic and clinical characteristics of study subjects.

	SCA3 (N = 74)		HC (N = 45)		SCA3 versus HC
	Mean	SD	Mean	SD	p‐valuea
Gender, F/M	49/25	NA	18/27	NA	0.495
Age, years	43.97	10.54	43.17	15.30	0.852
Height, cm	164.06	7.77	165.56	8.44	0.251
Weight, kg	59.34	10.88	61.7	10.78	0.175
BMI	21.86	3.04	23.09	2.96	0.164
Age at onset, years	37.11	9.98	NA	NA	NA
Disease duration	7.38	4.38	NA	NA	NA
Progression	1.59	1.05	NA	NA	NA
CAG	20.24	6.96	NA	NA	NA
SARA	8.96	2.49	NA	NA	NA
ICARS	24.11	7.18	NA	NA	NA
SARAEO	2.07	0.62	NA	NA	NA
ICARSEO	4.6	1.72	NA	NA	NA
ICARSEC	2.5	1.12	NA	NA	NA
Severity stage b
Mild	N = 17	NA	NA	NA	NA
Moderate	N = 57	NA	NA	NA	NA
Severe	N = 0	NA	NA	NA	NA

Abbreviations: CAG, expansion of cytosine‐adenineguanine; F, female; HC, healthy control; ICARS, International Cooperative Ataxia Rating Scale; ICARSEC, eyes closed subscore of ICARS; ICARSEO, eyes open subscore of ICARS; M, male; N, number; NA, not applicable; SARA, Scale for the Assessment and Rating of Ataxia; SARAEO, eyes open subscore of SARA; SCA3, spinocerebellar ataxia type 3; SD, standard deviation.

^aMann–Whitney U test.

^bMild: 3−7 on the SARA score; Moderate: 8−14 on the SARA score; Severe: more than 14 on the SARA score.

3.2. Analysis of Balance Variables to Discriminate SCA3 From HC

We conducted an analysis of the collected COP data and identified 52 potential balance variables derived from previous studies, literature sources [28], and clinical feasibility assessments to distinguish the balance capabilities of patients with SCA3 from HC subjects. The variables are expressed as mean or standard deviation (SD). As detailed in Table S2, sensor‐derived variables indicated that individuals in the SCA group exhibited inferior balance performance compared to the control group during both EC and EO standing task assessments. Specifically, the SCA group demonstrated greater speed and acceleration of bipedal COP position changes, as well as an increased total sway area relative to the control group (Figure 2A). Furthermore, the disparity in these variables between the SCA group and HCs was more pronounced in the EC task than in the EO task.

FIGURE 2.

Identification of highly discriminative center of pressure (COP)‐derived variables for distinguishing spinocerebellar ataxia type 3 (SCA3) patients from healthy controls (HCs). (A) Representative distribution of COP positions during a 3‐min eyes‐closed standing task. (B) Area under the curve (AUC) values of the 52 COP‐derived variables. (C) Radar chart showing the accuracy, sensitivity, and specificity of the 20 COP‐derived variables with AUC values greater than 0.9.

The Kruskal‒Wallis test revealed that all 52 variables exhibited statistically significant differences (p < 0.001) in their ability to differentiate between SCA3 patients and HCs. Notably, 42 of these variables demonstrated a large effect size (0.8 ≤ Cohen's d < 1.3), while eight variables exhibited an exceptionally high discriminative ability (Cohen's d ≥ 1.3). In addition to an extensive exploratory analysis of the 52 variables, their discriminative ability was also quantified by means of a receiver operating characteristic (ROC) curve and the calculation of the AUC (Table S3). Figure 2B shows the AUC values of some variables, sorted in descending order. It is worth noting that the AUC values for the top seven variables are all ≥0.95, demonstrating their substantial discriminative power in differentiating SCA3 patients from HC [29]. Specifically, they include the eyes closed, right side sway velocity (EC‐R Velocity, AUC = 0.97; Cohen's d = 1.638), eyes closed, right side sway acceleration (EC‐R Acceleration, AUC = 0.96; Cohen's d = 1.565), EC‐R Velocity SD (AUC = 0.97; Cohen's d = 1.622), EC‐R Acceleration SD (AUC = 0.96; Cohen's d = 1.528), eyes closed, right side sway velocity in antero‐posterior direction (EC‐R‐AP Velocity, AUC = 0.95; Cohen's d = 1.389), eyes closed, right side sway acceleration in antero‐posterior direction (EC‐R‐AP Acceleration, AUC = 0.95; Cohen's d = 1.258), and EC‐R‐AP Velocity SD (AUC = 0.95; Cohen's d = 1.411). The robust performance of these variables, as indicated by both Cohen's d and AUC values, underscores their efficacy in distinguishing between the groups. Furthermore, some additional variables exhibit AUC values greater than 0.9. The accuracy, sensitivity, and specificity of these variables are detailed in Table S4 and illustrated in the radar plot (Figure 2C). With the exception of certain variables exhibiting individual accuracy, sensitivity, and specificity indicators below 0.8, the majority of variables exhibit these three indicators exceeding 0.8. Notably, some variables, such as the EC‐R Velocity SD (Accuracy = 0.917; Sensitivity = 0.933; Specificity = 0.907; Cohen's d = 0.788), EC‐R‐AP Velocity SD (Accuracy = 0.917; Sensitivity = 0.911; Specificity = 0.920; Cohen's d = 0.794), EC‐R Acceleration (Accuracy = 0.925; Sensitivity = 0.933; Specificity = 0.920; Cohen's d = 1.149), EC‐R Acceleration SD (Accuracy = 0.907; Sensitivity = 0.933; Specificity = 0.920; Cohen's d = 1.158), and EC‐R‐AP Acceleration SD (Accuracy = 0.932; Sensitivity = 0.933; Specificity = 0.907; Cohen's d = 1.051), even surpass a threshold of 0.9. These variables are mostly derived from the COP measurements during the EC standing task, which possess substantial discriminative power, enabling precise differentiation between SCA3 and HC, and consequently reducing the probability of missed diagnoses and misdiagnoses.

Furthermore, a subgroup analysis was conducted to evaluate the system's sensitivity in early‐stage SCA3. When comparing mild patients (SARA score 3–7) with HCs, key variables such as the EC‐R Velocity SD and EC‐R Velocity maintained high discriminative power, with AUC values of 0.919 and 0.914, respectively (adjusted p < 0.001). This demonstrates the potential of these digital biomarkers for early clinical identification even before significant functional deterioration occurs (Table S8).

3.3. Correlation of SCA3 Balance Variables With Clinical Scales

The Kruskal–Wallis test and Pearson correlation coefficient were utilized to examine the correlations between each balance variable and the clinical scores of SARA and ICARS. Table S5 shows the statistical significance of the correlation between each balance variable of the patients with SCA 3 and their respective SARA and ICARS scores. To control for multiple comparisons, p‐values were adjusted using the FDR method. It is important to highlight that the correlations between these balance variables and the scores on the two scales exhibit inconsistency. Overall, each variable demonstrates a stronger correlation with SARA scores than with the ICARS scores. For example, the eyes open, left side sway velocity in the medio‐lateral direction (EO‐L‐ML Velocity) shows a moderate linear correlation with SARA (r = 0.491, adjusted p < 0.001), but a weak linear correlation with ICARS (r = 0.397, adjusted p < 0.001). Furthermore, a larger set of variables shows moderate correlations with SARA scores in comparison to ICARS scores. Specifically, 28 variables demonstrate moderately correlated with SARA scores (eyes open, left side sway velocity [EO‐L Velocity] SD, r = 0.534; eyes open, left side sway acceleration [EO‐L Acceleration] SD, r = 0.520; eyes open, left side sway velocity in antero‐posterior direction [EO‐L‐AP Velocity] SD, r = 0.515; eyes open, left side sway acceleration in medio‐lateral direction [EO‐L‐ML Acceleration], r = 0.515; EC‐R Acceleration SD, r = 0.453; eyes open, left side sway acceleration in antero‐posterior direction [EO‐L‐AP Acceleration], r = 0.453; EC‐R Velocity SD, r = 0.452; EC‐R‐AP Acceleration SD, r = 0.446; eyes closed, left side sway velocity [EC‐L Velocity] SD, r = 0.432…), whereas only five variables are moderately correlated with ICARS scores: EC‐L Velocity SD, r = 0.438; eye closed, left side sway acceleration (EC‐L Acceleration) SD, r = 0.408; EC‐R Acceleration SD, r = 0.406; eyes closed, left side sway velocity in antero‐posterior direction (EC‐L‐AP Velocity) SD, r = 0.405; EC‐R Velocity SD, r = 0.403. These variables are derived from the COP measurement during EC standing and are all expressed in SD, indicating that variables of variability have a more substantial relevance to clinical assessment.

Currently, there is no ataxia assessment scale specifically proposed for the clinical symptoms of SCA3, and the clinical evaluation of disease severity typically relies on the combined results of two scales. The five aforementioned variables, which demonstrate moderate positive correlations with the ICARS scores, also exhibit similar moderate positive correlations with the SARA scores (adjusted p < 0.001), as illustrated in Figure 3. Furthermore, these variables effectively discriminate between SCA3 patients and healthy individuals. Consequently, we further investigated the correlations of these five variables with scale subscores, severity stage, disease duration, expansion of CGA, AAO, and age, and the detailed statistical significance of their correlation are presented in Tables S6 and S7, where p‐values were corrected using the FDR method.

FIGURE 3.

Pearson correlations of high‐discriminative variables with clinical Scale for the Assessment and Rating of Ataxia (SARA) scores and the International Cooperative Ataxia Rating Scale (ICARS) scores.

Both ICARS and SARA include specific assessment items designed to evaluate balance function, where ICARS specifically contains both EO and EC posture assessment items, from which corresponding subscores ICARS_EO and ICARS_EC are derived and can be used independently to assess the patient's balance coordination ability. Considering that the five discriminant variables identified in our study were all derived from the EC standing task during the digital assessment, our analysis focused on examining the statistical correlation between these variables and ICARS_EC. As illustrated in Figure 4A and Figure S3A−C, three variables—EC‐L Velocity SD (r = 0.465, adjusted p < 0.001), EC‐L‐AP Velocity SD (r = 0.445, adjusted p < 0.001), and EC‐L Acceleration SD (r = 0.443, adjusted p < 0.001)—are moderately correlated with ICARS_EC. In contrast, the remaining variables demonstrate a weaker correlation with ICARS_EC, specifically EC‐R Acceleration SD (r = 0.373, adjusted p < 0.001) and EC‐R Velocity SD (r = 0.361, adjusted p < 0.001). Additionally, these variables are also weakly correlated with disease stage, as depicted in Figure 4B and Figure S3D−F. Notably, EC‐L Velocity SD (r = 0.566, adjusted p < 0.001) and EC‐R Acceleration SD (r = 0.545, adjusted p < 0.001) exhibited moderate correlations with disease duration. The other variables were weakly correlated with disease duration (Figure 4D and Figure S3G−I), and all five variables showed weak correlations (r < 0.4) with disease stage, CGA, and AAO (Figure 4D,E and Figure S4). However, several variables show a moderate correlation with age, as shown in Table S7. Specifically, all five variables (EC‐L Velocity SD, EC‐L‐AP Velocity SD, EC‐R Velocity SD, EC‐L Acceleration SD, and EC‐R Acceleration SD) maintained statistical significance (adjusted p < 0.05) in relation to age.

FIGURE 4.

Pearson correlations of some high‐discriminative variables with (A) ICARS eyes‐closed (ICARS_EC) scores, (B) disease duration, (C) disease severity, (D) expansion of cytosine‐adenineguanine (CGA), (E) age at onset (AAO), and (F) patient age.

3.4. SHAP‐Based SCA3 Discrimination

A comprehensive understanding of the severity of SCA3 is essential for developing an appropriate treatment strategy. However, the statistical analyses presented indicate that the selected feature variables exhibit only a weak correlation with disease severity, thereby limiting their utility in clinical decision‐making. Building upon the results of the above statistical analysis, we employed XGBoost in conjunction with SHAP values to conduct an interpretable analysis of SCA3 disease severity. In this study, the XGBoost model was trained on 70% of a randomly selected dataset, with the remaining 30% allocated for model validation. Due to the unequal distribution of mild and moderate patient data within the sample, the synthetic minority oversampling technique was employed during the data loading stage. This technique effectively alleviated the issue of class imbalance and enhanced the model's capacity to identify minority classes [30]. Prior to the training of the model, data dimensionality reduction was conducted using principal component analysis, which retained over 90% of the variance information and compressed the data to the optimal principal components, markedly improving computational efficiency.

The XGBoost classifier achieved an accuracy of 87% on the test dataset. Table 2 shows the model's performance for the category of mild and moderate patients, with 82% precision and 91% recall for mild patients, and 92% precision and 83% recall for moderate patients. The F1 scores were 86% and 87% for mild and moderate patients, respectively. These results demonstrate that the XGBoost model based on COP features effectively distinguishes between mild and moderate patients.

TABLE 2.

Performance metrics of the XGBoost model for classifying mild and moderate SCA3 patients.

Class	Precision	Recall	F1‐score
Mild patients	0.82	0.91	0.86
Moderate patients	0.92	0.83	0.87
Accuracy	/	/	0.87

Note: Mild patients: patients with a SARA score between 3 and 7; Moderate patients: patients with a SARA score between 8 and 14.

The SHAP summary plot in Figure 5A illustrates the relative importance of each parameter for the classification model. The columns represent the features of the dataset, and the rows represent the absolute values of the SHAP values. The results demonstrate that several statistically significant variables, including EC‐L Velocity SD (ǀSHAPǀ = 0.8989), EC‐L‐AP Velocity SD (ǀSHAPǀ = 0.8662), and EC‐L Acceleration SD (ǀSHAPǀ = 0.8989), exert the most substantial average impact on the model output magnitude on a global scale, thereby corroborating the validity of the preceding statistical evaluation. Conversely, the other two statistically significant variables, namely, EC‐R Velocity SD and EC‐R Acceleration SD, demonstrate minimal contributions to the model output magnitude, with SHAP values of 0.5845 and 0.5787, respectively. Meanwhile, the variables EC‐L‐AP Acceleration SD (ǀSHAPǀ = 0.8523) and EC‐L Acceleration (ǀSHAPǀ = 0.8061), despite their relatively weak statistical correlation with disease severity stage (r = 0.294, presented in Table S9), exhibit a substantial impact on the model output magnitude. This finding underscores the critical importance of dynamic features in complementing prior statistical analyses, thereby suggesting that multiple factors contribute to the disease's severity. Furthermore, it demonstrates that the severity of SCA3 can be accurately and effectively differentiated based on the COP variables, thereby enhancing the efficiency of clinical disease diagnosis.

FIGURE 5.

(A) Bar plot showing SHAP values, highlighting the top features contributing to the classification model's predictions. (B, C) SHAP dependency graphs illustrating the relationships of (B) eye closed, left side sway velocity in antero‐posterior direction (EC‐L‐AP Velocity) SD and (C) eyes closed, right side sway acceleration (EC‐R Acceleration) SD with SHAP values. The color gradient reflects the values of associated features.

To enhance our understanding of the model's ability to preidentify features, we have constructed the SHAP dependency graph, illustrated in Figure 5B,C. The color gradient in the scatter plots indicates the effect of the interaction features (variables depicted on the right axis) on the contribution of the target features (variables depicted on the bottom axis). The association between features of EC‐L‐AP Velocity SD and EC‐R‐AP Acceleration SD (Figure 5B) reveals a significant variation in SHAP values as the EC‐L‐AP Velocity SD progressively increases from 0 to 4000. Specifically, when the feature value of EC‐L‐AP Velocity SD falls within the range of 0−1000, the SHAP values are predominantly distributed in the negative range, indicating that the lower interaction feature values tend to drive the model toward classifying the sample as indicative of mild SCA3. As the value of EC‐L‐AP Velocity SD increases, the SHAP values become predominantly distributed in the positive range and exhibit larger magnitudes, indicating that higher feature values are associated with the model's inclination to classify the sample as indicative of moderate SCA3. Besides, the alteration in the contribution of the target feature of EC‐L‐AP Velocity SD is more pronounced when the value of the interaction feature of EC‐R‐AP Acceleration SD is elevated, suggesting that the interaction feature amplifies the prediction of the target feature at this juncture. Furthermore, a similar dependency trend was observed between the interaction feature of EC‐L Velocity SD and the target feature of EC‐R Acceleration SD, as depicted in Figure 5C. When the EC‐L Velocity SD value is low, the corresponding SHAP value is also low, indicating that lower interaction feature values tend to drive the model to classify the sample as indicative of mild SCA3. As the EC‐L Velocity SD increases, the SHAP value exhibits an overall upward trend, indicating that greater velocity fluctuation in the right foot increases the likelihood of the model classifying the patient as having moderate SCA3. Simultaneously, the target feature EC‐R Acceleration SD also increases, revealing a significant positive interaction between the two features. This interaction amplifies the model's ability to discriminate disease severity. Therefore, although EC‐R Acceleration SD does not rank highly in terms of its overall contribution to the model, its statistical significance, along with its interaction with other features in the dependency plot, suggests that it remains a crucial indicator for elucidating the dynamic performance of balance function in SCA3 patients. The SHAP analysis results indicate that the primary characteristics of standing balance not only independently affect the model's predictive outcomes but also elucidate the underlying neurological dysfunction that impairs dynamic balance control through an interactive mechanism. This enhances the model's reliability in identifying clinical conditions, reduces the costs associated with clinical assessments, and significantly accelerates the diagnostic process. These findings provide a critical reference for understanding the pathological features of SCA3 patients and for developing personalized treatment strategies.

4. Discussion

This study builds on our previously validated wearable plantar pressure insole platform and well‐characterized patient cohort to conduct a targeted, in‐depth analysis of static COP‐derived balance biomarkers for SCA3 assessment. We systematically evaluated the discriminative capacity of COP metrics to distinguish SCA3 patients from age‐matched HC, quantified their correlations with standard clinical ataxia scales, and developed an interpretable XGBoost‐SHAP framework to stratify mild and moderate SCA3 while uncovering previously underappreciated feature interactions. Unlike prior multimodal cross‐subtype classification efforts, this work focuses exclusively on static postural features to address a critical unmet clinical need: objective, quantitative severity stratification for SCA3, a gap that has persisted despite advances in wearable‐based differential diagnosis.

Over the past decade, wearable sensors [31], posturography [32], and machine learning [33] have been widely applied in research on hereditary cerebellar ataxias. Traditional laboratory force plates and optical tracking systems have confirmed that COP sway velocity, area, and directional displacement are sensitive indicators of postural instability in SCA3 [34], but these devices are limited to laboratory use. Portable wearable insoles break the spatial constraints and have become a mainstream tool for large‐sample clinical assessment [35]. Most existing wearable‐based SCA3 studies fall into two categories: distinguishing SCA3 patients from healthy individuals using gait or balance features, or combining multimodal data to differentiate SCA3 from other ataxia subtypes. Our prior work represented a typical cross‐subtype classification study: it integrated gait and static COP features to differentiate SCA3 and MSA‐C, two clinically similar cerebellar ataxias, and proved that EC‐state COP metrics and gait rhythm variability were core subtype‐specific biomarkers [22]. However, that study did not further explore whether these validated COP biomarkers could be used for within‐disease severity stratification, nor did it analyze the quantitative links between COP variables and clinical rating scales, disease duration, or age at onset. Additionally, feature interactions, which are critical for understanding complex motor dysfunction, were not investigated in the cross‐subtype classification task. The present study was specifically designed to fill these three interconnected research gaps.

Our results confirmed that COP variables collected during EC standing outperform EO metrics for differentiating SCA3 patients from HC, which aligns with the consensus in posturography research on cerebellar ataxia [36]. Cerebellar degeneration in SCA3 impairs proprioceptive and vestibular sensory integration, so patients rely heavily on visual input to maintain balance; removing visual feedback (EC condition) amplifies postural sway and enlarges group differences. The top‐ranked COP variables (e.g., EC‐R Velocity SD and EC‐R Acceleration SD) achieved AUC values above 0.95, and remained highly discriminative even when only mild SCA3 patients (SARA score 3–7) were compared with HC. This finding is particularly clinically significant, as it demonstrates that static COP biomarkers can detect subtle preclinical and early symptomatic balance deficits before they become apparent on traditional subjective rating scales. Notably, most high‐performance features were derived from the right foot in this cohort, which may relate to limb dominance and asymmetric weight‐bearing during quiet standing. This phenomenon requires further verification in larger cohorts with foot dominance recorded in future work.

To further evaluate the system's sensitivity in early‐stage identification, we performed a supplementary subgroup analysis comparing mild SCA3 patients with HCs (Table S8). The results confirmed that key variables, such as EC‐R Velocity SD, maintained high discriminative power even in this early symptomatic phase (AUC = 0.919, adjusted p < 0.001). This consistent performance in both the primary cohort and the mild subgroup underscores the sensitivity of these digital biomarkers to subtle postural instabilities. These variables demonstrate high sensitivity and accuracy, particularly during EC standing tasks without visual stabilization, where they reveal more pronounced differences between SCA3 patients and controls compared to EO conditions (Table S4). Velocity‐related COP variables emerged as the most sensitive parameters, with SCA3 patients exhibiting markedly increased COP velocity. This observation is consistent with prior reported evidence: gait velocity has been shown to correlate with disease severity in ataxia [37], and COP velocity was identified as one of the most discriminative postural stability indicators across visual conditions in platform‐based assessments [38]. The velocity of the COP is typically quantified by the mean integrated velocity, as well as the mean velocity in the ML direction and in the AP direction. In this study, we observed that highly discriminative variables were predominantly associated with the AP direction, where elevated values were observed compared with the HC. These observations are consistent with the findings of Salavati et al. [39], who reported increased swing amplitude and velocity in the AP direction under both eye‐open and eye‐closed conditions. This phenomenon may be attributed to deficits in lower‐limb motor control, such as low back pain or anterior cruciate ligament injuries, which necessitate frequent adjustments of the center of gravity to maintain balance. Additionally, larger envelope areas indicate more dispersed pressure trajectories, reflecting impaired balance control. This provides further evidence that mild‐to‐moderate SCA3 patients exert greater effort to stabilize their center of gravity during balance maintenance.

COP variables, like EC‐L and EC‐R Velocity SD, showed strong positive correlations (r > 0.4, adjusted p < 0.001) with SARA and ICARS scores, indicating their clinical relevance across the mild‐to‐moderate spectrum. Notably, these correlations remained statistically significant after FDR correction, underscoring the robustness of these digital measures. Besides, these variables also confirm that increased postural control variability reflects more severe motor dysfunction and disease severity, thus objectively complementing traditional subjective assessments. Similarly, the velocity of the COP variables showed significant correlations with patient characteristics (Table S6), with dynamic variables such as EC‐R Velocity SD and EC‐R Acceleration SD being strongly correlated (0.5 < r < 0.8) with disease progression and moderately correlated (0.3 < r < 0.5) with disease severity. These correlations highlight increased postural instability in patients with mild‐to‐moderate SCA3, who require more frequent and irregular motor adjustments to maintain balance. In addition to their diagnostic utility, variables of COP were significantly correlated with age and AAO (adjusted p < 0.05), as illustrated in Figure 4 and Table S7. These findings suggest that digital variables not only capture current motor dysfunction but may also reflect the cumulative burden of the disease and impairments specific to different stages. Moreover, numerical variables based on COP showed significant correlations with clinical ataxia scores, disease progression, CAG repeat number, and age at disease onset. These associations emphasize the capability of COP variables to comprehensively capture motor dysfunction in SCA3 and highlight their potential to complement clinical assessments as objective digital biomarkers. Overall, variables derived from COP quantify motor dysfunction in a multidimensional manner, effectively bridging the gap between subjective clinical observations and objective variables of disease severity. Accurately assessing the severity of ataxia holds substantial importance for formulating effective clinical treatment and rehabilitation strategies. Since limitations of traditional scales, including individual patient variability, varying quality of scale assessments, and inherent subjective factors, pose significant challenges to precise disease severity evaluation, our low‐cost wearable system provides a more accessible and objective alternative for potential longitudinal monitoring and digital performance measurement. The statistical analysis revealed that data obtained under the EC condition exhibited a stronger correlation with clinical severity compared to data collected under the EO condition with visual stabilization.

On the basis of statistical analysis, we constructed an XGBoost model for mild/moderate SCA3 stratification, which reached 87% overall accuracy. Combined with SHAP interpretability analysis, we identified EC‐L Velocity SD, EC‐L‐AP Velocity SD, and EC‐L Acceleration SD as the most influential features for severity classification. More importantly, SHAP dependency plots revealed significant interactive effects between COP features: the contribution of a single feature to severity prediction was amplified when paired with another elevated COP metric. This finding cannot be observed via conventional univariate statistical analysis, and provides a new perspective to understand the complex dynamic balance dysfunction in SCA3. Compared with single‐feature assessment, the combined effect of multiple COP variables can more comprehensively reflect the overall impairment of the postural control network.

The XGBoost model achieved a classification accuracy of 87%, with F1‐scores of 0.86 and 0.87 for patients with mild and moderate SCA3, respectively, underscoring the robustness of COP‐derived variables in distinguishing disease severity within the nonsevere spectrum. SHAP analyses further complemented these results (Figure 5) by identifying key features within the classification model and highlighting velocity variables under eyes‐closed conditions as the most influential variables (EC‐L Velocity SD, |SHAP value| = 0.899; EC‐L‐AP Velocity SD, |SHAP value| = 0.876; EC‐L Acceleration SD, |SHAP value| = 0.866). These variables also exhibited significant correlations in conventional statistical analyses, further reinforcing their clinical significance. In contrast, other variables, such as EC‐R Velocity SD (r = 0.566, adjusted p < 0.001) and EC‐R Acceleration SD (r = 0.545, adjusted p < 0.001), exhibited high correlations with the clinical disease process but had lower SHAP values, despite showing superior discriminatory effects in statistical analyses. This discrepancy suggests that SHAP analyses complement conventional methods in a multivariate setting by filtering redundant information arising from multicollinearity among COP variables and elucidating complex feature interactions and their influence on classification outcomes.

The EC‐L‐AP Velocity SD was identified as one of the most influential variables in the SHAP summary plot. However, as shown in Figure 5B, its influence does not follow a linearly increasing trend. The SHAP value for this parameter initially rose sharply with its own value, then plateaued and stabilized at a high level after reaching a medium‐high range. This pattern suggests that declines in patient balance are closely linked to significant changes in this parameter, potentially amplified by synergistic interactions with other variables as the disease advances. The contribution of EC‐L‐AP Velocity SD to disease severity was further enhanced by elevated EC‐R‐AP Acceleration SD values, as reflected by the color variations in the plot. This phenomenon indicates potential synergistic effects among multiple variables that collectively exacerbate balance deterioration as the disease progresses. These findings offer insights into the complex mechanisms underlying postural instability and highlight the importance of interactions among key indicators in assessing disease severity.

Similarly, for variables that do not exhibit prominent dominance in the SHAP summary plot, dependency plots revealed their interactions within specific ranges or under particular conditions. As illustrated in Figure 5C, the SHAP values for the EC‐R Acceleration SD increase in parallel with its elevation, exhibiting significant variability across different ranges and a pronounced nonlinear enhancement at higher values. This complexity suggests that EC‐R Acceleration SD not only shows strong discriminatory power in statistical analyses but also enhances its explanatory value in machine learning by dynamically interacting with other equilibrium variables. This complexity suggests that EC‐R Acceleration SD not only shows strong discriminatory power in statistical analyses but also contributes to model interpretation through its interaction with other COP variables. These findings provide a refined perspective on how COP‐derived variables reflect disease severity and highlight the value of SHAP analysis for interpreting feature interactions in the classification model.

In summary, this study extends the application of wearable‐derived COP balance biomarkers from cross‐subtype differential diagnosis to single‐disease severity assessment for SCA3. Static COP variables, especially metrics acquired under eyes‐closed standing, are robust digital indicators to distinguish SCA3 from healthy people, correlate well with clinical rating scales, and can be used for interpretable severity stratification. The objective quantitative framework established in this work can compensate for the shortcomings of subjective clinical scales. Future research will expand multicenter cohorts, carry out longitudinal monitoring, and integrate gait and postural features to develop an all‐in‐one wearable assessment system for the diagnosis, severity evaluation, and progression tracking of SCA3 and other cerebellar ataxias.

Nevertheless, several limitations should be acknowledged. First, SCA3 is characterized by genetic anticipation, whereby disease onset tends to occur at an earlier age in successive generations due to the instability and expansion of CAG repeats in the ATXN3 gene [40]. This genetic feature contributes to considerable heterogeneity in disease onset and progression across individuals. In addition, because the present study adopted a cross‐sectional design, longitudinal observations were not available, and thus the proposed digital biomarkers were not evaluated for their ability to predict disease progression or their sensitivity to longitudinal disease changes. Although longitudinal comparisons were beyond the scope of the current study, our previous work has demonstrated the feasibility of digital gait and balance biomarkers in cross‐disease classification tasks, particularly in distinguishing SCA3 from other neurological disorders in clinically challenging differential diagnostic scenarios. Future studies should incorporate prospective longitudinal follow‐ups and larger cohorts to further determine whether COP‐derived variables can sensitively capture disease progression and support personalized disease monitoring and intervention strategies.

5. Conclusion

Building on prior research of wearable digital biomarkers for ataxia differential diagnosis, this study further demonstrates that digital COP‐based variables effectively distinguish SCA3 patients from HC, and have potential as a noninvasive method for classifying disease severity. Furthermore, the continuous monitoring of these measures could serve as a valuable tool for assessing disease progression and evaluating treatment efficacy. This approach offers a reliable and interpretable alternative to traditional clinical assessments, enabling precise evaluation of disease severity in SCA3 through digital metrics. Future research should aim to validate and refine this method with larger patient cohorts and multimodal data integration, thereby paving the way for personalized management and precision interventions for SCA3.

Author Contributions

Y.Z. and X.‐Y.C.: Conceptualization, methodology, validation, data curation, writing – review and editing, project administration. Y.Z., S.‐R.G., and X.‐Y.C.: Funding acquisition. Y.X. and W.Z.: Methodology and investigation, formal analysis, writing – original draft. K.‐L.L., W.L., X.L., Q.‐K.S., R.‐Y.Y., J.N., and S.‐R.G.: Methodology, data curation, validation .

Conflicts of Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Ethics Statement

The study was conducted in accordance with the ethical principles posited in the Declaration of Helsinki–Ethical Principles for Medical Research Involving Human Subjects. The protocol was approved by the institutional review board. Participants underwent the consent process and provided documentation of informed consent prior to any study procedures. There was no participant compensation.

Supporting information

Supplementary Material: nyas70398‐sup‐0001‐SuppMat.docx