This study was approved by the Ethics Committee of Beijing Tiantan Hospital. Written informed consent was obtained from all the participants. A total of 248 participants diagnosed with PD (mean age: 63.46 ± 10.54 years) were recruited from Beijing Tiantan Hospital, Capital Medical University. Participants met the diagnostic criteria for PD established by the Movement Disorder Society (MDS) (Postuma et al., 2015). The exclusion criteria were as follows: (1) a history of stroke and cerebrovascular disease, (2) orthopedic impairment or other disease which may lead to gait disturbance, (3) MDS-UPDRS III 3.10: gait is score 4, (4) cognitive disorder was evaluated using Mini-mental State Examination (MMSE) and the cutoff values of MMSE for exclusion were adjusted by the education level where <18 for illiterate level, <21 for elementary level, and <24 for middle or above level (Katzman et al., 1988). Of these, video recordings of MDS-UPDRS III gait and posture-related items (#3.9–#3.13) were available for 225 participants and used for additional multi-rater scoring.

The MMSE and MDS-UPDRS III were administered and scored by a movement disorder specialist. An Motor Function and Motor Symptom Quantitative Assessment System (GYENNO SCIENCE, Shenzhen, China) (GYENNO Technologies Co. Ltd., 2022) was used in this assessment. This wearable motion and gait quantification assessment system is approved by Conformitè Europëenne Medical (CE Medical), National Medical Products administration (NMPA), and U.S. Food and Drug Administration (FDA). Moreover, this assessment platform has also supported research efforts at the intersection of medicine and engineering (Cai et al., 2023; He et al., 2024; Zhang et al., 2024; Lin et al., 2023). Participants performed a standardized gait assessment consisting of three consecutive trials, referred to as shuttle walk tests. Each trial required participants to walk straight along a 3.6-meter path, to execute a 180-degree turn, and return to the start position, while wearing ten inertial measurement unit (IMU) sensors (Figure 1). Two sensors were secured to the dorsal side of each wrist. The chest sensor was positioned over the sternum, while the waist sensor was placed at the level of the fifth lumbar vertebra (L5). For the lower limbs, a pair of thigh sensors were attached bilaterally, 7 cm above the knee, and a pair of shank sensors were positioned 7 cm below the knee joints. In addition, two foot sensors were fixed on the dorsal side of the metatarsus (instep) of each foot. All sensors were fastened firmly at their respective positions using adjustable straps. The 3.6-meter distance was selected due to its widespread use in Parkinson’s disease gait assessments, offering an optimal balance between patient safety, spatial feasibility in clinical environments, and its demonstrated ability to effectively provoke early gait abnormalities and freezing episodes (Stebbins et al., 2013; Choi et al., 2020).

For 225 participants with available video recordings of gait and posture-related items (#3.9–#3.13) during the MDS-UPDRS III assessment, a multi-rater, multi-round adjudication process was employed to minimize subjectivity and inter-rater variability. Two qualified raters independently assessed each video. For items where ratings differed, a consensus meeting was held after 1 week to re-evaluate and discuss. If consensus matched either original rating, it was adopted as the final score. If not, a third, more senior movement disorder specialist conducted an independent assessment, which was used as the final rating. Our analysis was conducted based on participants for whom video recordings were available, with the finalized re-rated scores from the adjudication process serving as the definitive ground truth labels.

Each IMU sensor consisted of 3-axial accelerometers and gyroscopes. Sensor data from the accelerometers and gyroscopes (x, y, and z-axis values) were continuously captured at a frequency of 100 Hz by the ten IMU sensors throughout the gait assessments in real time and were transmitted to the host computer via a Bluetooth link. The gait assessment was segmented into two types of phases: straight walking and turning, following the approach described in our previous studies (Lin et al., 2023). These phases were identified using kinematic signals from the waist, specifically the waist’s horizontal rotation angle. Distinct peaks in the waist rotation curve correspond to turning events. The first and second peaks indicate the onset and completion of the first turn, while the third and fourth peaks mark the start and end of the second turn. The remaining periods were classified as straight walking phases. Before feature extraction, the raw sensor data underwent a preprocessing pipeline. Specifically, the data were processed through a filtering procedure to reduce noise while preserving the true motion signals. Subsequently, orientation estimation was performed to convert the raw acceleration and angular velocity signals into meaningful spatial kinematic parameters (e.g., joint angles and angular velocities).

Based on the processed sensor data, a total of 240 kinematic features, such as Step Length, Cadence, and Double Support, were calculated. The definitions of these kinematic features are listed in Supplementary Table 1. Each participant completed three walking trials. For each trial, the walking assessment was segmented into two straight-walk sections and two turning sections, resulting in a total of six straight-walk sections and six turning sections per participant across all three trials, refer to straight-walk I (SW-1), 180° Turn I (T1), SW-2, T2, SW-3, T3, … and SW-6, T6. Extracted gait features were divided into three categories: (1) Segmentation-gait cycle-based features: these features were calculated within individual gait cycles but only for straight-walk sections. Within each straight-walk section, gait features were calculated for each gait cycle and then averaged across all gait cycles within that section. Averaging across all gait cycles in each section helps to minimize the impact of stride-to-stride variability and random fluctuations, providing a more stable and reliable estimate of each gait parameter for that section. By analyzing each section separately rather than only across the entire assessment, we were able to capture detailed, gait characteristics, which may be clinically relevant but could be masked if only whole-trial averages were considered. Gait cycles were detected by initial contact (IC) and terminal contact (TC) events as described in our previous studies (Lin et al., 2023). The right gait cycle begins from the right IC, then right TC, and then the right IC. The left gait cycle begins from the left IC, then the left TC, and then the left IC. Salarian et al. (2004) demonstrated that shank gyroscope signals are particularly effective for identifying IC and TC events during walking. Specifically, the first local minimum before and after each peak in the shank’s angular velocity was used to determine the timing of IC and TC events. Following a similar approach, we extracted IC and TC events from our data. For example, Trunk—Max Sagittal Angular Velocity, which was the measurement of the sagittal projection of the torso’s maximum angular velocity, it was calculated within individual gait cycles, and the values were then averaged across all gait cycles within a straight-walk section. (2) Segmentation features: this type of gait feature was calculated directly within each section, independent of gait cycle detection. These features reflect the overall performance or characteristics of an entire section (e.g., a straight-walk or turning segment) without relying on identifying precise gait events such as IC or TC. By considering the specific conditions of each section, segmentation features allow for a more comprehensive assessment of gait and posture characteristics during different phases of the assessment, providing complementary information beyond gait cycle-based features. For example, SW—Lumbar—Difference of Sagittal Sway was the difference of the sagittal projection of the waist’s tilt relative to the gravity vertical between the start and end moment of a straight walk section. (3) Whole assessment features: this type of gait features was calculated across the entire gait assessment, providing a global summary of gait performance. For example, WT – Trunk – Difference of Coronal Sway represents the difference in the trunk’s coronal tilt relative to the gravity vertical between the start and end of the whole trial. This approach captures overall changes or trends in gait characteristics throughout the entire assessment, allowing for a comprehensive evaluation of general gait stability and performance.

To represent overall gait characteristics, account for differences across straight-walk sections and across turning sections, and minimize bias due to limb dominance, feature construction was performed in two steps. Step1: For both segmentation-gait cycle-based features and segmentation features, the maximum, minimum, mean, and mean of difference across the six straight-walk sections and six turning sections were calculated. As a result, for each parameter, four summary features were generated, noted as—max, —min, —mean, and —diff_mean, respectively. For example, after feature construction step 1, feature Gait Speed L was replaced by Gait Speed L—max, Gait Speed L—min, Gait Speed L—mean, and Gait Speed L—diff_mean. Step 2: To quantify asymmetry and overall condition, for each pair of left- and right-side parameters, the maximum, minimum, and absolute difference were calculated, resulting in three representative features: (min), (max), and (diff). For example, after feature construction step 2, the pair of left- and right-side parameters Gait Speed L—mean and Gait Speed R—mean were replaced by Gait Speed—mean (max), Gait Speed—mean (min), and Gait Speed—mean (diff). This feature construction process was applied based on the features obtained from all three walking trials, ensuring that the final set of representative features for each participant integrated information from all trials and all relevant sections. This strategy enhances the stability and robustness of the extracted gait features while preserving section-level and side-specific information.

The primary demographic characteristics of all participants (n = 248) and re-rating sample (n = 225) are summarized in Table 1. There were no statistically significant differences observed in demographic variables between the overall cohort and the training and testing subsets, for both the original and re-rating samples. Figure 2 presents frequency histogram of scores for each gait and posture item. As described in the Methods section, merged categories (scores 0, 1, 2/3/4) were applied for items #3.9 (arising from chair) and #3.11 (freezing of gait), merged categories (scores 0, 1, 2/3) were applied for items #3.10 (gait), while merged categories (scores 0, 1, 2, 3/4) were applied for items #3.13 (posture). Figure 3 illustrates the distribution of scores on the MDS-UPDRS III gait and posture subscale, indicating fewer participants with higher scores.

Predictive performance comparisons between XGBoost and SVM algorithms for each of the five gait and posture items based on LOOCV are presented in Table 2. XGBoost consistently demonstrated better performance compared to SVM; therefore, XGBoost was selected as the final predictive model for all five gait and posture items. Weighted F1 score of item #3.9 ~ #3.11 is above 0.7, while they are approximately 0.6 for item #3.12 and #3.13. Macro-F1 score of item #3.9 ~ #3.11 is above 0.6, while they are approximately 0.5 for item #3.12 and #3.13. Absolute accuracy of all five XGBoost models exceeded 55%, with items #3.9, #3.10, and #3.11 achieving values above 70%. Acceptable accuracy (within ±1 point of true scores) exceeded 80% for all models, with notably high accuracy (>90%) observed for items #3.9, and #3.10. Kw coefficients, indicating the level of agreement between true and predicted scores, were above 0.5, representing at least moderate agreement for all five models. Detailed precision, recall, and F1 values for each gait and posture item classification by XGBoost models are shown in Table 3. Overall, precision and recall varied across score categories and items, with lower score categories (indicating less severe impairment) generally showing higher classification performance, while higher score categories, particularly those with limited sample sizes, exhibited reduced performance. For item # 3.9, the model achieved strong classification performance for the unimpaired category (Score 0), with LOOCV precision, recall, and F1 score of 0.855, 0.803, and 0.828, respectively. While performance for higher severity categories showed room for improvement, these categories were notably underrepresented (e.g., only 7 samples for Score 2/3/4 in LOOCV), which likely contributed to reduced model performance in those groups. In item # 3.10, the model demonstrated good performance for Score 2/3 with a precision of 0.916 and recall of 0.784 under LOOCV, while classification for Score 0 showed lower performance, with an F1 score of 0.48. The model also demonstrated robust performance for item # 3.11, achieving an F1 score of 0.889 for the non-impaired group (Score 0) and 0.627 for higher severity categories (Score 2/3/4) under LOOCV, suggesting reliable detection of both absence and presence of freezing phenomena, despite limited data for intermediate severity levels. Performance for items # 3.12 and # 3.13 followed similar trends, with higher classification metrics for lower severity categories and reduced performance for higher scores, primarily attributable to class imbalance and the small number of samples representing more severe symptoms. Confusion matrices for the training datasets of the final XGBoost model are presented in Figure 4.

Feature importance (Gain) utilized in the final predictive models for items #3.9 to #3.13 is detailed in Supplementary Tables 2–6. The total number of features selected for predictive models of items #3.9, #3.10, #3.11, #3.12, and #3.13 was 35, 35, 30, 30, and 40, respectively. Importantly, several key features demonstrated meaningful clinical correlations. For example, Shank—Swing RoM—max (max) had a negative correlation (R = −0.476, p < 0.001, Figure 5A) with scores on item #3.9. Higher scores on item #3.9 indicate more severe gait and posture impairment, which is consistent with reduced lower-limb mobility. A smaller shank swing range may reflect impaired lower-limb strength and coordination, affecting functional tasks such as standing up from a chair. Effective Trial Duration (R = 0.522, p < 0.001) had a positive correlation, while Shank—Swing RoM—mean (max) (R = −0.629, p < 0.001) had a negative correlation with scores on item #3.10 (Figures 5B,C), which were consistent with clinical findings. As the more severe the gait impairment is, the slower the walking speed will be and the smaller the range of motion of the shank will be. 180° Turn—Steps—mean had a positive correlation (R = 0.482, p < 0.001, Figure 5D) with scores on item #3.11. Higher scores on item #3.11 indicate more severe gait and posture deficits, which is consistent with increased step count during turning. Increased mean step count during turning may indicate gait freezing tendencies or impaired postural control in PD patients. 180° Turn—Max Angular Velocity—max (R = −0.586, p < 0.001) had a negative correlation, while 180° Turn—Duration—mean (R = 0.604, p < 0.001) and Straight-Walking Duration—mean (R = 0.551, p < 0.001) had positive correlation with scores on item #3.12 (Figures 5E–G), which were consistent with clinical observations. As the more unsteady the participant is, the slower they walk, the more time they would spend on walking. 180° Turn -Total Duration—mean had a positive correlation (R = 0.54, p < 0.001), while Trunk—Forward Sway Max—max had a negative correlation (R = −0.461, p < 0.001) with scores on item #3.13 (Figures 5H,I), which were also consistent with clinical findings. The more sever posture impairment, the more time the participant spends on turning. Trunk—Forward Sway Max—max was used to describe the sagittal projection of the trunk’s maximum forward tilt relative to the gravity vertical while walking (backward: positive value, forward: negative value). The larger the absolute value of the negative value, the greater the participant’s trunk forward angle while walking, the severer the posture damage, the higher scores on item #3.13.

Predicting the MDS-UPDRS III gait and posture subscale score via LOOCV on the training data achieved a MAE of 1.349 and a RMSE of 1.645. The correlation between predicted and actual subscale scores was strong (R = 0.798).

Sensor contributions to predictive models for each gait and posture item are presented in Supplementary Table 7. Because some features incorporated multiple sensors, the cumulative sensor contributions exceeded 100% for some gait and posture items. Shank sensors provided the greatest contributions for gait and posture items #3.10 (48.6%), #3.11 (43.3%), #3.12 (43.3%), and #3.13 (35%). For item #3.9, the chest sensor contributed most (37.1%), followed by the shank sensor (31.4%) and lumbar (28.6%). Comparison of the contributions of different sensors revealed that shank sensors provided the greatest contribution, followed by chest sensor.

The full dataset of 225 participants was divided into training (80%) and independent test sets (20%). Performance evaluations of the final predictive models on the independent test set are provided in Tables 2, 3. Test data performance closely matched training LOOCV results, confirming the models’ reproducibility capability. Weighted F1, macro-F1, absolute, and acceptable accuracy values for items #3.9, #3.10, and #3.11 remained high (above 0.7, 0.67, 70, and 90%, respectively). However, predictive performance for items #3.12 and #3.13 was suboptimal. Table 3 shows class-specific evaluations on test data. The predictive model for item #3.9 showed high precision and recall for Score 0 and Score 1 (above 70%). No participants were predicted to be Score 2/3/4, which lead to the precision value to be NA. Model on gait and posture item #3.10 had high precision and recall value on Score 0 and Score 2/3 (above 75%). The fraction of instances correctly classified as Score 1 out of all instances the model predicted to belong to Score 1 was 46.2% (precision), while it was 50% for recall. Model on gait and posture item #3.11 exhibited good performance for Score 0 and Score 2/3/4 but limited ability for Scores 1. Confusion matrices for the test dataset of the final XGBoost model are presented in Figure 6. Furthermore, the predictive model for gait and posture subscale was evaluated on independent test data, which achieved a MAE of 1.432, a RMSE of 1.776, and a strong correlation coefficient (R = 0.818) between the predicted gait and posture subscale and true gait and posture subscale.