In this study, 20 participants with idiopathic PD were enrolled. All patients met the 2015 Movement Disorder Society (MDS) diagnosis criteria for primary PD (16), including 8 males and 12 females, the patients were between stage 1 and 3 on the H&Y scale (namely early-to-middle stage patients). At the same time, 17 age- and gender-matched healthy participants were recruited as the healthy control (HC) group, including 9 males and 8 females. Participants from both groups were between 45 and 85 years old, and all of them could walk 10 meters or more without assistance. The participants with significant systemic diseases (such as musculoskeletal, cerebrovascular, cardiovascular, respiratory) and other neurological diseases or uncorrected visual disturbances or ailments that might modify the gait patterns were actively selected and excluded from the study. This study was approved by the Ethics Committee of Central Hospital of Dalian University of Technology (Reference No. YN2022-039-57). All the study participants signed the informed consent document prior to the study participation. The study was performed according to the guidelines of the declaration of Helsinki.

Demographic information was collected, including age, gender, height, weight and disease duration. General neurological examination and clinical scales assessment were performed and recorded by two experienced neurologists in movement disorders. Dopaminergic therapies were recorded, levodopa equivalent morning dose was calculated (levodopa equivalent dose (LED) is calculated as levodopa dose + levodopa dose × 1/3 (if on entacapone) + piribedil (mg) + pramipexole (mg) × 100 + selegiline (mg) × 10 + rasagiline (mg) × 100 + amantadine (mg) + controlled-release levodopa (mg) × 0.75) (17).

The severity of motor symptoms in PD patients were assessed using the H&Y staging scale and MDS-UPDRS III scale. H&Y staging scale ranges from 0 (no symptoms) to 5 (wheelchair bound or bedridden unless aided), MDS-UPDRS III: each item ranges from 0 to 4 (0 = normal, 1 = slight, 2 = mild, 3 = moderate, 4 = severe), there are 33 items in total, and the score of each item are added up to get a total score, ranging from 0 to 132 points.

The gait data acquisition system consists of a self-developed handheld device and two sensor nodes (sampling frequency is 200 Hz) (Figure 1). Each sensor node is integrated with a microcontroller, a WIFI wireless communication module, a 450 mAh lithium battery and an inertial measurement unit (IMU) sensor (specific parameters are shown in Table 1), all those are packaged in a customized 3D printing box. The size of sensor node is 4.6 × 3.4 × 2.2 cm and the weight is 42 g. Each sensor node is attached to the lateral ankle of each foot (Figure 2). The working state of the sensor nodes is controlled by the handheld device. As “Start” is clicked, the two sensor nodes will start to collect data synchronously. The participant will stand for 3–5 s firstly (the data is used for correction at this time) and then start to walk. When the “Stop” command is sent to the sensor nodes, both nodes will be suspended, the collected data will be uploaded through wireless network and gait analysis will be performed.

Proprietary data acquisition technology is used to preprocess the original data and extract gait variables, calculate plantar dorsiflexion angle (PDA), plantar flexion angle (PFA), stride length, foot clearance, velocity, cadence and stride time (specific definitions are shown in Table 2 and Figure 3). The accuracy of the gait data acquisition system have been validated in previous studies (18), compared to the gold standard (optical motion capture), the position estimation error is less than 1% with regard to three dimensional motion.

The gait acquisition was performed in an obstacle free and flat environment. The participants were asked to walk independently through a 10-meter-long sidewalk at self-selected comfortable speed, in order to assess gait parameters objectively. There were no repeating instructions during the walk to avoid auditory cues. The first and last steps were excluded. Adequate protective measures were taken during the walking to ensure that the participants would not fall down suddenly.

PD patients were evaluated at both OFF state and ON state. In the OFF state, PD patients were fasting overnight, and at least 72 h without using dopamine agonists and at least 12 h without using any anti-Parkinson drugs. MDS-UPDRS III was scored and gait was assessed by two experienced neurologists in movement disorders. After the OFF test, they were re-tested in the ON state by the same neurologists, which was 2 h after 1.5 times of the regular morning levodopa dose or levodopa/benserazide (Modopar) 100/25–200/50 mg in the drug-naive patients. Healthy controls completed the same gait assessment but did not receive levodopa.

In this study, we used gait parameters as input features and scale scores as output labels to train machine learning regression models. Specifically, we employed linear regression (implemented via Scikit-learn) as the modeling algorithm, which is suitable for small-sample regression tasks due to its simplicity and interpretability. To account for scale differences among features and improve model stability, all input features were standardized using z-score transformation prior to model training. The model was trained using default hyperparameters (regularization strength C = 1.0, solver = ‘liblinear’).

To further investigate the predictive value of gait parameters for the tremor and non-tremor components of the patient’s score, we divided the output labels into three categories: (1) MDS-UPDRS III score, (2) Tremor part score (i.e., items 3.15 + 3.16 + 3.17 + 3.18), and (3) Non-tremor part score (i.e., MDS-UPDRS III score minus tremor part score). A total of three prediction models were constructed.

Leave-one-out cross-validation was used to evaluate the performance of each model to assess the generalization ability and prediction accuracy of the model. Leave-one-out cross-validation is an effective method for evaluating the performance of regression models on a given data set. It does this by using one sample as the validation set and the remaining samples as the training set in each iteration, repeating this process until each sample has been used as the validation set, and finally averaging all the validation results to evaluate the performance of the model.

During the evaluation, the following metrics were used: R-squared (R²), Mean Absolute Error (MAE), and Mean Absolute Percentage Error (MAPE). R² measures the goodness of fit of the model to the data, with a range from 0 to 1, and the closer it is to 1, the better the model fits the data. MAE measures the average absolute error between the predicted value and the true value, and the smaller values correspond to higher accuracy of the model. MAPE, on the other hand, is the average absolute percentage error expressed as a percentage, which can better assess the relative error between the predicted value and the true value, and is often used to understand the prediction accuracy of the model in different ranges.

The statistical analysis was carried out using SPSS 26.0 (IBM, Armonk, NY, United States) in our study. The Shapiro–Wilk test was used to check the normality of the data. Continuous variables with normal distributions were presented as means ± standard deviations (x¯±s), the Independent Samples t-Test was used to analyze the differences between continuous variables. Continuous variables with non-normal distributions were presented as medians and interquartile distances [M (P₂₅, P₇₅)], and the Mann–Whitney U Test was used to analyze the differences. The Chi-square test was used to compare categorical variables. The Paired Samples t-Test or Wilcoxon Matched-Pairs Signed-Rank Test was used to analyze the changes between OFF and ON states. The Pearson correlation (r) or Spearman’s rank correlation (r_s) was used to analyze the correlations between MDS-UPDRS III score and gait parameters. The statistically significant difference was considered p < 0.05 in two-tailed tests. In this study, machine learning regression models were trained and scatter plots were generated using Python 3.8 with sklearn and matplotlib.

The demographic information of PD group and healthy control group was shown in the first four lines of Table 3. We collected the data of gender, age, height and weight of 20 PD patients and 17 healthy controls. There was no statistically significant difference in their demographic characteristics, and p > 0.05 for all items was in line with the principle of object selection.

We collected the disease duration, H&Y stage and MDS-UPDRS III score in the OFF stage of PD patients; MMSE score in the ON stage of PD patients. The disease duration in PD group ranged from 1 to 8 years (median was 2.5 years), MMSE score ranged from 20 to 30 points (median was 27 points), H&Y stage ranged from 1 to 3, the mean MDS-UPDRS III score (OFF) was 36.55 ± 15.31, as shown in Table 3.

The comparisons of gait parameters between the PD patients (OFF state) and the healthy controls are shown in Table 4 and Figure 4. Various gait parameters were significantly lower in PD patients than in healthy controls, including PDA (left, right), PFA (left, right), stride length (left, right), foot clearance (left), velocity (p < 0.05), the p values were lower than or equal to 0.001 in PDA, stride length and velocity among the above parameters, which indicated that PD patients tended to have a flat-footed gait pattern, slow velocity and short stride length when walking. We speculated that the statistically significant difference in foot clearance between left and right was due to the different statistical methods we used, the Independent Samples t-Test was used to analyze the differences between left foot, while and the Mann–Whitney U Test was used to analyze the differences between right foot, which leaded to the decrease of efficiency.

In this study, we analyzed the correlation between gait parameters and MDS-UPDRS III score in PD patients in OFF state. MDS-UPDRS III score was negatively correlated with PDA (right), PFA (left, right), stride length (left, right), foot clearance (left, right) and velocity. These correlations indicated that the increased severity of motor symptoms was associated with the more obvious of the gait disorders, such as the flat-footed gait pattern, slow velocity, short stride length and low foot height. The PFA had a stronger correlation with the severity of motor symptoms compared with PDA. (Table 5 and Figure 5).

We also explored the effect of levodopa on gait parameters. The differences between OFF and ON states were shown in Table 6 and Figure 6, MDS-UPDRS III score was decreased in the ON state, PDA (right), stride length (left, right), foot clearance (right), velocity, cadence were increased in the ON state, stride time was decreased in the ON state, and the differences were statistically significant (p < 0.05), highlighting the therapeutic effects of levodopa on gait parameters and motor symptoms in PD patients. In our study, only the PDA and foot clearance on the right side had statistically significant difference, we speculated that it might be related to the unilateral onset of PD. Among the 20 patients, 3 had left side symptoms, 6 had right side symptoms, and 11 had bilateral symptoms, which might be related to more severe symptoms on the right side.

It could be seen that the model had the strongest explanatory ability to the non-tremor part score (R² = 0. 775) from the evaluation indexes shown in Table 7, the MDS-UPDRS III score could also be interpreted by the model (R² = 0. 675), and the model was weak in interpreting the tremor part score (R² = 0. 138).

The scatter plot (Figure 7) showed the relationship between the predicted value and the true value of the regression models. From the plot, it could be seen that the model for predicting MDS-UPDRS III score and non-tremor part score performed better. At the same time, we constructed a baseline model using only stride length and velocity to predict MDS-UPDRS sub-scores. These models yielded lower R² values (e.g., stride length (left and right) model: R² = 0.493 for non-tremor part score, velocity-only model: R² = 0.467 for non-tremor part score), confirming the predictive value of incorporating multiple gait features. (Figure 8).

We further analyzed the feature importance of the gait parameters, feature importance was evaluated based on the values of the regression coefficients, the magnitude of each standardized coefficient reflects the strength of its association with the outcome (Figure 7). The importance degree of MDS-UPDRS III score was in the following order: PFA (right), PDA (right), cadence, PDA (left), stride time, PFA (left), stride length (right), foot clearance (right), velocity, stride length (left), foot clearance (left); the importance degree of non-tremor part score was in the following order: PDA (right), PDA (left), PFA (right), cadence, PFA (left), stride time, stride length (right), foot clearance (right), velocity, stride length (left), foot clearance (left). We could see that PDA and PFA contribute most to MDS-UPDRS III score and non-tremor part score. Each standardized coefficient of tremor part score was very small, we did not analyze it.