The data collection process in this study used a rigorous screening and matching strategy (Figure 1). The initial research cohort comprised 193 participants who were preliminarily categorized into 3 groups based on cognitive functional status: PD with normal cognition (PD-NC; n=54), PD with subjective cognitive decline (PD-SCD; n=36), and PD with mild cognitive impairment (PD-MCI; n=103). During the first-stage screening process, some participants were excluded due to incomplete demographic records, partial scale completion, or missing gait measurement data. Specifically, 21 cases were excluded from the PD-NC group, 9 cases from the PD-SCD group, and 17 cases from the PD-MCI group. After preliminary screening, the sample sizes for each group were 33 cases for PD-NC, 27 cases for PD-SCD, and 86 cases for PD-MCI. To control for the influence of age as a significant confounding factor on research outcomes, we used age-stratified propensity score matching for all participants. This matching process ensured balanced age distribution across the 3 groups, effectively controlling for age differences’ impact on cognitive function and gait performance. After age matching, the final sample size included in the analysis was 65 participants, consisting of 14 cases in the PD-NC group, 21 cases in the PD-SCD group, and 30 cases in the PD-MCI group. This matching strategy not only enhanced the internal validity of the study but also provided a more reliable data foundation for subsequent machine learning model analysis.

A consecutive inclusion of 65 age-matched patients from a movement disorder clinic was undertaken (PD-NC [n=14]; PD-SCD [n=21]; and PD-MCI [n=30]). Both physical and thorough interviews were done with them. Age, gender, education level, and duration of disease were recorded. All of the patients were diagnosed with idiopathic PD according to the International Parkinson and Movement Disorder Society (MDS) clinical diagnostic criteria. The inclusion criteria were as follows: (1) diagnosis of PD according to the criteria of the MDS, (2) stages 1-3 on the Hoehn-Yahr (H-Y) scale, (3) stable use of antiparkinsonian medication based on medical history, and (4) ability to walk independently for at least 5 minutes. Patients were excluded based on the following criteria: (1) those who diagnosed dementia, (2) a diagnose of secondary PD, (3) medical history indicates having history of brain disorders (eg, cerebral infarction, cerebral hemorrhage, or brain tumor), and (4) those with serious psychiatric diseases (eg, Hamilton Anxiety Scale [HAMA] ≥21 [30], Hamilton Depression Scale [HAMD] ≥17 [31], or obsessive-compulsive disorder). This study was approved by the Research Ethics Committee of the Second Affiliated Hospital of Nanchang University.

The participants underwent a comprehensive neurocognitive assessment. Motor symptom severity was evaluated using the MDS Unified Parkinson Disease Rating Scale (MDS-UPDRS) and H-Y stage [32]. Cognitive function was assessed using the MoCA. The MoCA covers multiple cognitive domains, including memory, language, complex visuospatial processing, and executive function, and has a maximum score of 30. Emotional features were assessed with the HAMA and HAMD, which has been validated for use with patients with PD.

According to the guidelines for the diagnosis of dementia in PD published by the MDS, a cut-off score of <26 is recommended for PD-MCI; additionally, 1 point is added to the total score (if less than 30) for participants with ≤12 years of education [33]. In this study, we define normal global cognition as MoCA ≥25/30 for participants with ≤12 years of education and 26/30 for participants with >12 years of education. Participants answered item 1.1 of MDS-UPDRS Part I, and based on the score, the participants were divided into 2 groups: a score of ≥1 indicated those with subjective cognitive complaints and a score of <1 indicated those without subjective cognitive complaints. Thus, we defined PD-SCD as individuals with item 1.1 of MDS-UPDRS Part I ≥1 as well as normal objective global cognition; PD-NC was identified as patients with PD with item 1.1 of MDS-UPDRS Part I <1 and normal objective global cognition [34,35].

Participants performed single-task TUG (TUGst) and dual-task TUG (TUGdt) gait trials wearing a Motor Function and Motor Symptom Quantitative Assessment System (MATRIX; GYENNO SCIENCE). The MATRIX consists of 10 inertial sensors (ie, 10 data recording channels) sampling at 100 Hz. Each inertial sensor consists of a (1) triaxial accelerometer with range=±16 g and sensitivity=16,384 LSB/g and a (2) triaxial gyroscope with range=±2000 dps and sensitivity=131 LSB/dps. The chest sensor was placed on the sternum of the chest, and the lumbar sensor was attached to the fifth lumbar vertebra. Two wrist sensors were bilaterally placed on the dorsal side of the wrist. Two thigh sensors were bilaterally placed 7 cm above the knee, while 2 shank sensors were bilaterally placed 7 cm below the knee. Two-foot sensors were bilaterally placed at the instep (dorsal side of the metatarsus) of each foot. All sensors were tightened to designated locations using straps. As a 3-axis accelerometer and a 3-axis gyro are combined in one sensor, each sensor of the MATRIX records 6 separate signals synchronously. Therefore, 60 signals (10 sensors×6 signals) were recorded in computers for feature extraction of a single TUG trial per participant. Please refer to previous research literature for the specific attached position and detailed metadata of sensors [36,37].

During the TUGst test, without the use of any mobility aids, participants stood up from a chair, walked 5 meters in a straight line, turned 180°, walked straight back to the chair, turned another 180°, and sat on the chair at their usual pace. During the TUGdt test, participants performed the usual-pace TUG test while subtracting 7 serially from 100 (ie, 93, 86, 79, and so on). To avoid learning effects and fatigue, participants performed both tests quietly and completed only one trial of each test.

To illustrate motion profiles of the arms, lumbar spine, trunk, feet, and shanks, 209 kinematic parameters were synthesized for individual TUG trials by gait event recognition based on signal data stored in computers [36,37]. Parameters were listed in Multimedia Appendix 1. This process involved several key aspects: data acquisition and preprocessing using MATRIX with 10 inertial sensors, each recording 6 independent signals (triaxial accelerometer and gyroscope), resulting in 60 raw signals per TUG trial; gait event recognition through algorithmic identification of key gait events, dividing the trial into walking phase, stand-to-sit transition, sit-to-stand transition, and turning phases; parameter extraction and classification based on gait event recognition, extracting 209 kinematic parameters from raw signals covering spatial and temporal characteristics such as velocity, range of motion, stride length, cadence, swing, and stance, as well as asymmetry features including velocity asymmetry and range of motion (RoM) asymmetry; and anatomical plane analysis considering sagittal, coronal, and transverse planes across multiple body segments including arms, lumbar spine, trunk, feet, and shanks. These parameters represent motion asymmetry for bilateral limbs, kinematic variability (SD of parameters), and task-related spatial and temporal characteristics, including velocity (degree per sec), range of motion (RoM, degree), peak velocity (degree per sec), and RoM (degree) of 3 anatomical planes, respectively (parameters the same as lumbar), stride velocity (m/s), cadence (step per min), gait cycle (sec), stride length (cm), step length (cm), swing (%), stance (%), peak angular velocity (degree per sec), RoM (degree), velocity asymmetry (%), RoM asymmetry (%), absolute deviation between limbs, walk duration (sec), and turning duration (sec). According to parameter definitions, 209 parameters were categorized into variability features, amplitude features, pace features, speed features, axial features, and asymmetry features of motion, ensuring comprehensive capture of motion characteristics across all body segments during TUG tasks and providing rich kinematic information for cognitive state classification. Please refer to previous research literature for specific algorithms [37].

Furthermore, DTC parameters, describing the magnitude of the effect of the cognitive challenge on motion performance, were calculated with equation

, in which i refers to each of the 209 parameters.

This study was approved by the Research Ethics Committee of the Second Affiliated Hospital of Nanchang University, with all participants providing informed consent. Data were anonymized using ID numbers, and follow-up participants received travel compensation.

There were no statistically significant differences among the PD-NC, PD-SCD, and PD-MCI groups in terms of demographic characteristics (including age, sex, height, thigh length, weight, education level, disease duration, and H-Y stage) or scores on MDS-UPDRS I, MDS-UPDRS II, MDS-UPDRS III, HAMA, and HAMD (P>.05).

The diagnostic criteria for PD-NC, PD-SCD, and PD-MCI were based on UPDRS1.1 scores and MoCA total scores, which showed statistically significant intergroup differences (P<.001; Multimedia Appendix 2). The MoCA scale and its subdomain scores exhibited different patterns of differences among the 3 groups (Figure 2). In terms of MoCA total scores, the PD-MCI group (23.00, 95% CI 22.00-25.00) was significantly lower than both the PD-NC group (28.00, 95% CI 27.00-29.00) and the PD-SCD group (27.00, 95% CI 27.00-28.00), while no significant difference was observed between PD-NC and PD-SCD groups. Among the MoCA subdomain scores, statistically significant differences were found among the 3 groups in abstraction (P=.002), delayed recall (P=.005), and visuospatial ability (P<.001), with the PD-MCI group showing significantly lower performance in these cognitive domains compared to PD-NC and PD-SCD groups. In attention, the differences among the 3 groups approached statistical significance (P=.05), with the PD-MCI group showing a trend toward decreased attention. However, no statistically significant differences were observed among the 3 groups in executive function (P=.25), language (P=.83), naming (P=.45), and orientation (P=.11; Figure 2). These findings align with the disease progression pattern of PD. As a chronic neurodegenerative condition, cognitive function may gradually deteriorate with disease progression, particularly in higher-order cognitive functions such as memory, visuospatial ability, and abstract thinking. The significant decline in delayed recall and visuospatial ability observed in patients with PD-MCI reflects the typical characteristics of PD-related cognitive impairment. In contrast, language, naming, and executive functions remain relatively preserved in the early stages, which is consistent with the domain-specific pattern of cognitive impairment in PD.

Screening procedures identified 45 kinematic parameters meeting the established criteria, distributed across TUGst (n=25), TUGdt (n=12), and DTC (n=8) paradigms. Gait phase analysis revealed 35 parameters from walking phases, 9 from stand-to-sit transitions, and 1 from sit-to-stand transitions. Feature type distribution demonstrated a predominance of variability features (n=20), followed by pace (n=12) and axial (n=8) characteristics. This distribution pattern reflects differential impacts of cognitive decline on gait control mechanisms, with movement precision (variability) and temporal coordination (pace) emerging as critical cognitive function indicators. TUGdt paradigm analysis revealed pronounced movement differences between PD-MCI and both PD-NC and PD-SCD groups, particularly in variability, amplitude, pace, and axial domains. Notable examples include “Shank-Max Sagittal Angular Velocity R Std” (21.12 vs 28.38 vs 21.38) and “Trunk-Left Rotation Max Std” (1.94 vs 1.99 vs 1.33) in variability measures. Amplitude analysis of “Shank-Symbolic Symmetry Index” demonstrated compensatory patterns (14.27 vs 9.75 vs 12.62), suggesting potential compensatory mechanisms during SCD stages. TUGst paradigm differences primarily manifested in pace characteristics, exemplified by “Double Support” (23.16 [4.76] vs 18.55 [3.04] vs 20.87 [3.75]) and “Stance _MAXLR” (62.17 [2.75] vs 60.06 [1.95] vs 61.17 [2.09]). DTC parameter analysis revealed a concentration of differences in variability features, including “Gait Speed R.Std” (–0.97 vs –0.30 vs –0.14) and “Shank-Max Sagittal Angular Velocity R.Std” (–1.67 vs –0.33 vs –0.30), demonstrating cognitive load effects on motor performance. Cross-paradigm analysis identified consistent significant differences in specific features, such as “Step Length_ABSLR” in DTC (–0.36 vs 0.16 vs 0.14), establishing step length asymmetry as a robust cognitive function biomarker. These findings provide objective kinematic biomarkers for early cognitive state identification in PD, with TUGdt parameters demonstrating superior discriminative capacity.

Multimedia Appendix 3 results demonstrate systematic kinematic feature differences across cognitive states, primarily in movement precision (variability) and temporal coordination (pace) domains. Enhanced differences in TUGdt paradigms align with theoretical frameworks predicting motor control degradation under increased cognitive load conditions. Compensatory pattern identification (eg, preserved performance in SCD stages) suggests early-stage compensatory mechanisms maintaining motor function. However, progressive cognitive decline appears to compromise these mechanisms, resulting in significant motor deterioration in MCI stages. These findings establish objective kinematic biomarkers for early cognitive state identification and monitoring in PD, with TUGdt variability features demonstrating particular clinical translational potential.

Figure 3 presents boxplots illustrating the distribution of 10 selected kinematic parameters across the PD-NC, PD-SCD, and PD-MCI cognitive groups, visually demonstrating the differences among these groups for each parameter. The box in each boxplot represents the 25th percentile (Q1, bottom line of the box), the median (central line of the box), and the 75th percentile (Q3, top line of the box). The bottom and top of the error bars (whiskers) indicate the “Minimum” (Q1-1.5×IQR) and “Maximum” (Q3+1.5×IQR), where IQR is the interquartile range. Blue dots outside this range represent outliers.

As observed from the figure, for several DTC related parameters, such as “DTC_Lumbar-Right_Left Sway Max,” “DTC_Stand To Sit-Duration,” “DTC_Trunk-Max Transverse Angular Velocity,” and “DTC_Trunk-Right Rotation Max,” the median values for the PD-NC group are generally lower than those for the PD-SCD and PD-MCI groups, indicating lower motor costs under dual-task conditions in the normal cognition group. This aligns with the theory that impaired cognitive function leads to decreased motor control ability. For instance, “DTC_Stand To Sit-Duration” shows a median of approximately –0.05 for the PD-NC group, while the medians for the PD-SCD and PD-MCI groups are significantly higher than 0.20, suggesting that patients with cognitive impairment are more affected by cognitive challenges during the stand-to-sit transition. In dTUG (dual-task TUG) -related parameters, such as “dTUG_Arm-Asymmetry Of Max Sagittal Angular Velocity SD” and “dTUG_Lumbar-Right Sway Max SD,” the median values for the PD-SCD group appear slightly lower than those for the PD-NC and PD-MCI groups. This might reflect a compensatory mechanism or different motor strategies used during the SCD stage. For sTUG (single-task TUG) related parameters, such as “sTUG_Cadence L SD,” the median for the PD-NC group (approximately 3.5) is notably lower than that for both PD-SCD and PD-MCI groups (both approximately 5.5), indicating lower gait variability in cognitively normal patients under single-task conditions, whereas patients with cognitive impairment exhibit higher gait variability. These visual differences are consistent with the statistically significant results reported in Multimedia Appendix 3, further supporting the potential of these kinematic parameters as biomarkers for distinguishing different cognitive states.

The best-performing model in the 3-class classification was the SVM model. Its classification accuracies on the training and test sets were 0.8 (0.69-0.86) and 0.6 (0.45-0.73), respectively, indicating a good performance on the training set but reduced performance on the test set (Multimedia Appendix 4). Regarding recognition capability per category on the training set, PD-MCI had a relatively higher recall (0.71), while the PD-NC and PD-SCD categories exhibited relatively higher precision, at 0.90 and 0.86, respectively. On the test set, PD-MCI maintained a relatively higher recall (0.78), while PD-NC and PD-SCD demonstrated relatively lower precision, at 0.75 and 0.29, respectively. Based on the receiver operating characteristic (ROC) curves (Figure 4), the macro-average ROC AUC on the training set was 0.87, compared to 0.60 on the test set.

Integrating participants’ cognitive levels allows for a deeper explanation of the mechanistic differences in classifying MCI, SCD, and NC (Multimedia Appendix 5, Figure 5, and Multimedia Appendix 6). Regarding the distinction between MCI and NC, cognitive level analysis revealed significant differences between the 2 groups across multiple cognitive dimensions: the MoCA total score of the MCI group was significantly lower than that of the NC group (effect size, Cohen d=2.672, indicating a large effect); delayed recall ability was significantly lower in the MCI group than in the NC group (Cohen d=1.915); visuospatial ability was significantly lower in the MCI group than in the NC group (Cohen d=1.354). These significant differences in objective cognitive function provide a cognitive load basis for TUG task features to effectively distinguish MCI from NC. Due to limited cognitive resources, patients with MCI exhibit significant degradation in motor performance during dual-task TUG. This cognitive-motor interaction pattern can be effectively captured by TUG features.

Regarding the distinction between SCD and NC, cognitive level analysis revealed the fundamental cause of confusion. The SCD and NC groups showed no significant difference in MoCA total score (P=.14). Except for abstract ability, no significant differences were found in other cognitive subdomains, with most comparisons showing small-to-medium effect sizes.

This similarity in objective cognitive function explains why TUG features struggle to distinguish SCD from NC: although patients with SCD subjectively perceive cognitive changes, their objective cognitive resources are sufficient, leading to similar performance in dual-task TUG as NC, resulting in TUG features failing to effectively distinguish these 2 groups.

Concerning the distinction between MCI and SCD, cognitive level analysis showed significant differences in objective cognitive function between the 2 groups: the MoCA total score of the MCI group was significantly lower than that of the SCD group (Cohen d=2.301, large effect); delayed recall ability was significantly lower in the MCI group than in the SCD group (Cohen d=1.423); visuospatial ability was significantly lower in the MCI group than in the SCD group (Cohen d=1.528). However, despite significant cognitive differences between MCI and SCD, TUG features still faced difficulties distinguishing them, with 42.9% of SCD samples misclassified as MCI.

Feature importance analysis of cognitive levels further validated the cognitive level analysis results. The MoCA total score had the highest discriminative power (F_2,16=18.94; P<.001), followed by delayed recall (F_2,16=11.70; P<.001). This aligns highly with the significant decline in delayed recall observed in the MCI group in the cognitive level analysis (effect size, Cohen d=1.915; Figure 5). The decline in delayed recall ability may directly impact patient performance in dual-task TUG, as this task requires simultaneously processing motor and cognitive demands; the efficiency of cognitive resource allocation directly affects motor performance.

Furthermore, to gain a deeper understanding of the model’s decision-making mechanism, we used 2 feature importance analysis methods (Figure 6). Permutation importance analysis evaluated the impact of features on model performance by randomly shuffling feature values, revealing that DTC_Trunk-Right Rotation Max (importance value=0.145) was the most critical feature, followed by DTC_Trunk-Max Transverse Angular Velocity (0.090) and dTUG_Lumbar-Right Sway Max Std (0.080). These findings emphasize the importance of trunk motor control under dual-task conditions for distinguishing cognitive states. SHAP analysis provided 3 different perspectives on feature importance assessment. The SHAP feature importance bar plot showed that sTUG_Cadence L Std (mean [SHAP]=0.23) had the highest average impact, followed by DTC_Trunk-related features. This discrepancy with permutation importance results reflects the different mechanisms of the 2 analysis methods; permutation importance focuses on the impact of feature absence on overall model performance, while SHAP analysis focuses on the direct impact of feature value changes on model output. The SHAP feature importance summary plot further revealed the complex relationship between feature values and their impact on model output. For most DTC features, higher feature values tended to produce larger positive SHAP values, pushing the model output toward higher values, indicating that increased trunk kinematic parameters under dual-task conditions are associated with cognitive decline. This pattern supports the cognitive-motor interaction theory, suggesting that when cognitive resources are limited, individuals exhibit greater motor variability under dual-task conditions. SHAP waterfall plots provided detailed explanations at the individual sample level, demonstrating how each feature drives predictions from baseline values to final outputs. For example, in correctly classified PD-MCI samples, DTC_Trunk-Max Transverse Angular Velocity and DTC_Trunk-Right Rotation Max contributed most strongly to predictions, while correctly classified PD-NC samples exhibited relatively stable dual-task motor features. These individual-level analyses revealed the model’s decision-making mechanism in distinguishing different cognitive states, particularly the challenges faced when processing groups with similar cognitive states (such as PD-SCD and PD-NC; Multimedia Appendix 7).

In this study, we included 65 patients with PD exhibiting various cognitive states (PD-NC: 14 cases, PD-SCD: 21 cases, and PD-MCI: 30 cases) and explored the method of using kinematic characteristics for the stratified cognitive stage classification of patients with PD by combining wearable sensors with machine learning. We first described the gait characteristics of PD-NC, PD-SCD, and PD-MCI, respectively; second, we explored the factors associated with gait kinematic parameters and cognitive status; next, we analyzed significantly different gait parameters among groups; and finally, we established a model that relies on various TUG task outcomes, which can distinguish individuals with PD-NC, PD-SCD, and PD-MCI. Our results demonstrated that kinematic features are associated with cognitive stages, patients with PD-SCD have underlying kinetic impairments similar to those with PD-MCI, and the diagnostic SVM model showed potential in differentiating cognitive impairments. The SVM model achieved classification accuracies of 0.8 (0.69-0.86) and 0.6 (0.45-0.73) on the training and test sets, respectively, with PD-MCI demonstrating the highest recall rates (0.71 on the training set, 0.78 on the test set) and macro-average ROC AUC values of 0.87 and 0.60 for the training and test sets.

Regarding the classification performance, the model demonstrated effective differentiation between MCI and NC groups due to significant differences in objective cognitive function (MoCA total score effect size, Cohen d=2.672), which provided a cognitive load basis for TUG task features to effectively distinguish these groups. However, the model faced significant challenges in distinguishing PD-SCD from both PD-NC and PD-MCI groups. Concerning the distinction between SCD and NC, cognitive level analysis revealed the fundamental cause of confusion. By definition, SCD refers to a subjective perception of cognitive decline, which is not necessarily aligned with the actual cognitive level. For instance, substantial cognitive impairment may occur without being perceived, whereas minor declines may be perceived as significant. Thus, it is difficult to determine which condition reflects a higher level of cognition. Consequently, NC and SCD may not exhibit clear differences in cognitive impairment; the 2 groups showed no significant difference in MoCA total score (P=.14). It is correspondingly challenging to distinguish between them based on motor performance. This suggests that SCD may represent a subtype with a higher risk of progressing to MCI, rather than a group that is necessarily closer to MCI in terms of cognitive function.

Regarding the distinction between SCD and MCI, cognitive level analysis showed significant differences in objective cognitive function between the 2 groups (MoCA total score effect size, Cohen d=2.301). However, despite significant cognitive differences between MCI and SCD, TUG features still faced difficulties distinguishing them, with 42.9% of SCD samples misclassified as MCI. This seemingly paradoxical result highlights an important mechanism of TUG task features: although MoCA assessment revealed significant cognitive differences between SCD and MCI, this difference primarily manifests in static cognitive tests. In the dynamic context of dual-task TUG, cognitive differences may be insufficient to produce noticeable variations in motor performance. Specifically, while patients with SCD had significantly higher MoCA scores than patients with MCI, their relative cognitive advantage during dual-task TUG might be counterbalanced by other factors, such as attentional allocation strategies, task priority adjustments, or motor control patterns. Furthermore, UPDRS score analysis revealed that the SCD group had the lowest UPDRS III (motor examination) score (20.143), indicating relatively better motor function in patients with SCD. This motor advantage might further compensate for the impact of cognitive differences on TUG performance. This underscores that performance in the TUG task, as a cognitive-motor dual-task paradigm, is not solely influenced by static cognitive function. Instead, it results from the combined effects of dynamic cognitive-motor interactions, attentional allocation, motor control, and other factors. When cognitive differences are counterbalanced by other factors in a dynamic task context, TUG feature-based classification models struggle to accurately distinguish different cognitive states.

Additionally, the limited sample size (n=65) after age-matching may have contributed to model performance instability, with the training set containing only 45 samples and the test set containing only 20 samples, including only 7 SCD samples in the test set, resulting in insufficient sample representativeness that hindered the model’s ability to adequately validate SCD classification performance.

There are several limitations in our study. First, the MDS-UPDRS Part I 1.1 score ≥1 evaluating SCD was too simple compared with a comprehensive screening scale. However, SCD identified by MDS-UPDRS Part I 1.1 has been revealed to be associated with the development of PD-MCI, which validates the clinical significance of our measurement tool. Second, we used a short MoCA scale to assess global cognition rather than a comprehensive neuropsychological test focused on each cognitive domain. It is necessary to confirm our results with comprehensive neuropsychological tests. Next, the classification performance of the model needs to be further improved. Furthermore, a larger sample and follow-up multiple-center studies are essential to validate our cross-sectional results.

Our study examines kinematic features to distinguish between 65 patients with PD, including PD-NC, PD-SCD, and PD-MCI. Forty-five gait parameters differed significantly by group in the TUGst, TUGdt tests and DTC, with TUGdt parameters showing stronger cognitive association. Patients with PD-MCI had worse gait performance than patients with PD-NC and PD-SCD. The PD-SCD group showed reduced gait speed and swing speed but longer steps and strides, and our multiclass SVM classification model with kinematic parameters achieved a recall rate above 0.7 in both training and validation datasets. This suggests patients with PD-SCD could have early kinetic signs of cognitive impairment, positioning them between PD-NC and PD-MCI, and kinematic parameters could serve as predictors to distinguish the cognitive stage of patients with PD.