In the main findings for people with PD and controls of this study, the five selected features that were most relevant for the classification of people with PD and controls were the inner step length, step width, inner double support phase, thorax ROM, and incline angle during the 360° turning task. The machine learning approach showed that RF solved the PD classification problem with 98.1% and 98.0% accuracies for PD_36 and PD_5, respectively. In the main findings for freezers and non-freezers, the six selected features that were most relevant for the classification of freezers and non-freezers were outer step length, inner hip and ankle ROM, total distance of the COM, maximum anti-phase, and outer contralateral temporal coordination parameter during the 360° turning task. The machine learning approach showed that RF had 79.4% accuracy for FOG_36 and LR had 72.9% accuracy for FOG_6. Additionally, the 360° turning characteristics such as outer contralateral temporal coordination parameter, maximum anti-phase, and outer step length were associated with the clinical characteristics of people with PD and freezers. Therefore, the 360° turning features based on full-body kinematic analysis may enable classification of people with PD and controls, freezers and non-freezers, and its association with clinical characteristics is demonstrated. These findings are discussed in detail below.
Classification using feature selection through stepwise regression
In our study on the classification of people with PD and controls, 5 features were selected through stepwise regression to obtain the sensitive cutoff values in the ROC analysis. These turning features are related to the spatiotemporal parameters and turning strategy for the inner step of the more affected limb. People with PD showed a significantly shorter step length, wider step width, longer double support phase, greater thorax ROM, and smaller incline angle for maintaining their center of gravity between the two feet when compared to those with controls [44]. The supplementary motor area, which receives input from the impaired basal ganglia in people with PD, participates in the control of postural coordination and affects the bilateral function of gait [45]. These results may cause dynamic instability during turning because people with PD present a lower supplementary motor area activity than with controls [46]. Therefore, we suggest that people with PD and controls may be distinguished using turning features such as spatiotemporal parameters, trunk ROM, and incline angle related to coupling between posture and gait during turning tasks for the inner step of the more affected limb [47].
The machine learning approach showed that RF resolved the PD classification problem with 98.1% and 98.0% accuracies for PD_36 and PD_5, respectively. From these results, the possibility of distinguishing between people with PD and controls based on the 360° turning characteristics was confirmed to some extent. In the PD classification problem, the feature selection approach by stepwise regression showed reasonable accuracy performance. RF outperformed all other classifiers with all 36 features; in addition, LR, SVM, and RF with the reduced feature set performed better than the other classifiers in resolving the classification problems. These results indicated that feature selection by stepwise regression removed irrelevant features. Generally, the output of a model can be affected by multiple features. When the number of features increases, the model becomes complicated. An overfitting model tends to consider all features, even though some of them have very limited effect on the final output [48].
For classification of freezers (disease duration: 8.39 ± 5.83 years) and non-freezers (disease duration: 4.36 ± 3.61 years), six features were selected via stepwise regression to obtain the sensitive cutoff values in the ROC analysis. These turning features are related to the turning strategy and interlimb coordination. Freezers showed a significantly shorter step length, greater hip ROM, smaller ankle ROM, longer total distance of the COM, smaller maximum anti-phase, and longer contralateral temporal coordination parameter using the compensatory strategy for postural instability when compared with those of non-freezers. In particular, to observe a phase delay between the upper and lower limbs in people with PD and freezers, temporal coordination while turning may be used as the primary parameter. During turning, delayed temporal coordination between the upper and lower limbs indicates a reduced coordination capacity [49, 50]. In addition, our result showed that freezers have a dependent turning characteristic by shortening the outer step length of the rotation center, along with en bloc head and trunk rotation compared to non-freezers [51]. These characteristics may increase the risk of falls owing to potential FOG characteristics, suggesting that they may experience greater turning difficulty due to increased postural instability with disease progression [52]. It may be caused by deficits in several components of postural control, such as anticipatory postural adjustments, delayed reaction time, abnormal automatic postural reactions, and abnormal axial kinesthesia [53]. The turning task threatens the stability of freezers more than any other freezing trigger as it requires a precise control of each limb [26]. In addition, freezers showed less rhythmic and uncoordinated gait patterns when compared to those of non-freezers [45]. These results suggest that freezers may experience difficulties in performing automatized movements without adequate attention [54] and may be more vulnerable to impairments related to interlimb coordination because turning is asymmetrical when compared with a straight gait [55]. Therefore, we suggest that freezers and non-freezers can be classified based on the turning features related to postural transitions and coordination [56].
The machine learning approach showed that RF resolved the FOG classification problem with 79.4% accuracy for FOG_36, and LR resolved it with 72.9% accuracy for FOG_6. From the results, the possibility of distinguishing between freezers and non-freezers based on the 360° turning characteristics was confirmed to some extent; however, the FOG classification problem appears more challenging than the PD classification problem. First, no classifier had high accuracy of more than 80%. Moreover, the SD of the accuracy for FOG_6 was higher for all classifiers except KNN and QDA when compared with the results for FOG_36; especially, the SDs of the accuracies of SVM and LR showed a rapid increase (the value for SVM ranged from 8.9 to 16.8% whereas that for LR ranged from 0.8 to 10.8%). We speculate that this was caused by the relatively small sample size in this study. The small number of samples might cause a misinterpretation in the mathematical optimization procedure while the classifier is being trained, and it might affect the performance of SVM and LR because of the nature of these classification algorithms. In future research to improve the accuracy of the FOG classification problem, the raw time series motion data during the 360° turning task need to be studied via advanced deep learning techniques such as the n-dimensional convolutional neural network and recurrent neural network. Although the raw motion data are converted to selected 36 features, there is a possibility of losing key information required to solve the FOG classification problem.
Associations between clinical and 360° turning characteristics of people with PD
This study conducted feature selection using stepwise regression for the 360° turning characteristics. Based on the selected turning characteristics, the associations between the clinical and turning characteristics of people with PD and freezers were investigated. We observed the associations between the clinical characteristics such as the UPDRS total and UPDRS III scores, Hoehn and Yahr stage, PIGD score, and NFOGQ score, and the selected features during the 360° turning task. Although our result was similar to the findings of the previous studies on the associations between the severity of PD and turning characteristics [57,58,59,60], most studies employed small sample sizes and often did not control for confounders that may affect the turning characteristics owing to physical characteristics such as age, sex, height, and BMI. In addition, the previous studies assessed the clinical characteristics in the “On” state of medication [58, 60], whereas this study assessed the clinical characteristics and turning task of people with PD in the “Off” state of medication. The medication status of people with PD influences the motor symptoms and may affect the generalization limitations of the associations between clinical and turning characteristics of people with PD who exhibit FOG [61, 62]. A previous study reported that people with PD and freezers showed a more constrained movement during turning in the “Off” state of medication when compared with the controls and non-freezers [63]. The study considered a compensation strategy for preventing falls in people with PD and freezers, which were caused by the declined ability to control the centrifugal forces that create the inertia forces to allow body rotation, especially immediately after the pivot point during turning [7]. In addition, as dynamic stability is already compromised in people with PD and freezers, they may have shown more careful movement during the turning [64, 65]. A more constrained postural strategy may be used to facilitate effective turning under dopamine depletion, which may influence the control of automatized movement [63, 66]. Especially, freezers need a strategy to increase their stability during turning owing to greater impairment of cognitive, executive, and attentional resources when compared with non-freezers [20, 67, 68].
We showed that PD severity for motor symptoms is related to a decrease in turning performance. Turning is an asymmetric task, in which one limb generates a stepping pattern, and the other helps with weight shifting and support; thus, it requires a higher level of bilateral coordination in people with PD and freezers [45]. In a majority of such people, the right limb is initially affected to a greater extent, suggesting a decline in the neural control ability during turning due to certain associations between the symptom-dominant side and dominant hemisphere [45]. In addition, a higher PIGD score was significantly associated with greater maximum anti-phase and shorter outer step length while turning in people with PD. This result showed that the severity of axial symptoms and gait difficulties during turning, and not the general severity of PD, might affect the turning performance [57]. Our study using the 360° turning task for the inner step of the more affected limb identified association with clinical characteristics of people with PD and freezers through the difference in turning characteristics according to disease severity for motor symptoms. Previous studies have reported that the more affected limb of people with PD tends to be affected predominantly throughout disease progression and may promote greater motor deficits [16, 69]. This suggests that the turning difficulty may be a result of asymmetry between the more and less affected limbs and impaired in both automatic and controlled processes [9, 70]. Therefore, we suggest that a more challenging 360° turning task for the inner step of the more affected limb may be evaluated through the turning performance of people with PD and freezers.
Furthermore, clinical characteristics related to PD severity, such as UPDRS total and III scores, PIGD score, Hoehn and Yahr stage, and NFOGQ score, were identified as the indicators of FOG [71]. Previous studies have shown the association of the severity of FOG with motor deficit [72, 73]. It has been suggested that induced motor deficit such as the loss of automaticity along with stepping inhibition during turning led to repeated weight shifts without stepping, resulting in trembling of limbs related to FOG [71, 72]. In particular, our result indicated that the outer step length decreased as the NFOGQ score increased in freezers. In this study, no difference between the inner and outer step lengths in freezers was observed during the turning task. These results do not indicate the asymmetry of steps during turning in freezers with advanced disease severity [74, 75], which may be reflected as reduced normal asymmetric gait strategy and bilateral motor coordination during turning [74].
Additionally, we observed the correlation of the NFOGQ, UPDRS total, PIGD score, and levodopa equivalent dose with disease duration. The advanced severity and long duration of the disease along with disease progression in people with PD may contribute to the severity of FOG [75]. There was also a significant correlation between the PIGD score and disease duration, which could lead to axial symptoms including gait disturbance and postural abnormalities in freezers with longer disease duration when compared with non-freezers [76]. Although people with PD are likely to develop FOG over time (it may be noted that all people with PD do not develop FOG), other factors such as the disease duration and dopaminergic treatment as well as genetic status may also influence gait disturbance [77].
Our study had several limitations. First, the effects of the “On” and “Off” states of medication and the differences in the turning direction were not compared while evaluating the 360° turning tasks. Second, our datasets have an imbalance related to gender and use different sample sizes. The results are expected to improve with a more homogeneous dataset. However, we analyzed after adjusting for the covariates of age, sex, height, and BMI. Third, the sample size of freezers in the FOG classification problem was relatively small: 34 freezers and 43 non-freezers. Although we used the random oversampling technique to handle this imbalanced dataset, the inadequate sample size was likely to cause instability of the classification performance of SVM and LR, as mentioned previously. In addition, a FOG episode was induced in one participant during turning for the inner step of the more affected limb; the corresponding results were excluded from the analysis. Fourth, the R2 values for the associations between disease severity and turning characteristics of people with PD are weak. Thus, further study using instruments to assess various clinical characteristics in the medication “On” and “Off” states and longitudinal studies are needed to generalize the associations between disease severity based on the clinical characteristics and the turning characteristics. Fifth, for previous studies, many measures related to disease severity (duration of disease, UPDRS total and III scores, and levodopa equivalent dose) of people with PD have been significantly different between freezers and non-freezers [31, 71, 72, 76]. However, our result that no difference in UPDRS III between freezers and non-freezers (p = 0.565). Research suggested that although UPDRS III may contribute to assessing the functional impact of FOG, there do not reflect the overall severity of FOG [31]. Therefore, further research is needed considering the sample size and objective evaluation status of freezers and non-freezers. Lastly, machine learning techniques with higher predictability for classification and a filtering technique for motor symptoms of people with PD and freezers need to be developed. A method employing a larger sample size or important factors contributing to improving the evaluation of disease severity and predictability of classification and diagnosis may be added to the classification model. Further studies are needed to evaluate the realistic patient movements on the raw time series motion data through advanced machine learning techniques such as deep learning.
The findings of this study have some important implications. First, the results of the turning characteristics for the inner step of the more affected limb in people with PD and freezers may be helpful in improving the clinical assessment and understanding of disease severity by disease progression. Second, the machine learning approach to resolve the PD and FOG classification problems of this study showed similar results when using kinematic features selected through 360° turning analysis. This result may be helpful in understanding the movement characteristics and classifying the disease severity of people with PD and freezers based on certain main factors of the spatiotemporal and kinematic features during turning tasks. Third, the clinical characteristics were shown to be associated with the turning characteristics. These results may help in ameliorating the motor symptoms of people with PD and improving the rehabilitative strategies, which may reduce the occurrence of freezing.