**6. Discussion**

The availability of low-cost home-based solutions for the reliable and automated assessment of motor symptoms in Parkinson's disease is highly desirable since it could provide several advantages, among which: reduction of costs and patient discomfort; better and prompt supervising and adjustment of the therapy; healthcare analytics for patient care improvements. Surely, among the features that these solutions should exhibit, particularly important are: a non-invasive approach to the assessment; a user-friendly interaction suitable to motor impaired users; an objective, continuous and automated evaluation of patient status, strongly correlated with the standard clinical assessments; an improvement of the reliability respect to the typical intra and inter-rater variability of the clinical evaluations.

In this paper, a self-managed system for the automated assessment of Parkinson's disease which tries to implement many of the aforementioned features is presented. The developed system is focused on posture instability and motor impairments of lower limbs and it is one of the elements of a larger project aimed to bring an overall automated assessment of UPDRS tasks at home [35].

As a first step, we addressed both the non-invasiveness and user-friendly interaction by a low-cost system based on an RGB-D optical device, then providing a gesture based human computer interface for the self-management of the assessment procedures. The usability of the interface was tested and verified by PD users during a campaign of data acquisition sessions. Then, the accuracy of the kinematic measures, as obtained by the system, was validated successfully by comparison with a gold standard equipment (i.e., an optoelectronic system). This was a necessary preliminary requirement, since an objective evaluation of the patient status is based on the strong correlation existing between motor impairments and kinematic parameters extracted from patient's movements.

To reliably refer the system assessments to the clinical ones, the analysis of possible movements was constrained to those specified by the UPDRS tasks. An experimental protocol was designed in which PD patients and healthy controls were assessed at the same time both by neurologists and by the system during the execution of the specific standard tasks defined by UPDRS. A feature selection procedure yielded to sets of optimal parameters, both correlated to UPDRS clinical scores and statistically significant in discriminating PD subjects from healthy controls. As shown in Figure 8, not all these parameters have the same discriminant power to separate subjects among the different PD severity classes; this is true especially for the AC and Po tasks. This is probably due to the limited number of PD subjects examined: consequently, further experiments could improve the current results.

Following related works based on wearable systems [40,41], the postural stability of PD subjects was characterized by CoM movements. We analyzed the CoM trajectories during the two phases of the Po task (named PSCOM), assuming the Phase2 as a mild secondary motor task [38]. As in [40,41], large differences in CoM trajectories of PD respect to HC were found. Differently to [41], a good correlation between PSCOM parameters and the standard postural stability test (PIGD) was observed. This result can be explained because of the different physical quantity and derived parameters considered by the two approaches: CoM displacements in our case, derivative of CoM accelerations in [41]. On the other hand, the CoM parameters in [40] have a closer physical relationship with ours: respect to us, the authors did not find a significant correlation between the PIGD scores and the parameters they selected, but this could probably be due to the exclusion of the retropulsion task from their analysis. In conclusion, we found that the PSCOM parameters are related to PIGD score and are also statistically significant: in fact, they clearly discriminate PD subjects from healthy controls, supporting the initial hypothesis of a worsening of PD stability during the execution of secondary tasks.

The automated assessment of UPDRS tasks is performed by means of kNN, MLR and SVM supervised classifiers, trained on the sets of selected parameters and the corresponding UPDRS scores from reference datasets of performances of PD and HC cohorts. In general, the accuracy of the SVM classifiers is better than those of the MLR and kNN classifiers. Besides, the binary-classification (i.e., HC versus PD) gives quite better results than the multiclass-classification, as expected. Moreover, in the last case, the classification error for the optimized SVM was never greater than 1 UPDRS class for all the tasks, and on the average well below of this value. This indicates that chosen classifiers are robust and, in any case, they do not make assessments too far from neurologists. Furthermore, these results agree with Table 10 about the measure of the inter-rater agreement ICCN12-SY, which indicate that the system performs almost as a third neurologist, except for PSPIGD task. For this task, the lower value of ICCN12-SY as compared to ICCN12 can be due to CoM parameters that are not directly comparable to PIGD subscale assessments or to the limited number of PD subjects included in the training set.

Due to the novelty of our approach, based on low-cost optical RGB-D device, we cannot compare directly the results of the classification accuracy with other similar works. Furthermore, a limited attention has been devoted to the automated assessment of specific UPDRS tasks by motion capture technologies. Then, we decided to refer to approaches based on wearable devices employing supervised classifiers [63]. Even if not directly comparable with our tasks, Timed Up and Go (TUG) test in [64] discriminate PD from HC by machine learning approach, with accuracy of 77.5%, which is lower than the value we have obtained (Table 11). In [6] the accuracy values for the multiclass classification of the LA and AC tasks are about of 43%, which are lower than ours (Table 11), even if care must be taken because the number of classes considered is different.

Summarizing, to our knowledge this is the first time that posture instability and lower limb motor tasks were assessed with reference to the clinical UPDRS context by a system based on optical RGB-D device. The results on the classifier accuracies and on ICC show that the automated assessments of the system are comparable with the clinical ones, then demonstrating their effectiveness. Furthermore, it is also the first time that a system based on low-cost optical device characterizes CoM movements for the assessment of Parkinson's Disease. Finally, another original feature is the interpretation of the posture improvement during quite stance as secondary motor task, and the findings about its effectiveness in assessing postural instability in PD subjects.

Certainly, some aspects of this work require a further investigation. For instance, the number of analyzed subjects should be increased to obtain a more robust characterization of each single task and a better accuracy in the automated assessments. Furthermore, the PD subjects should be distinguished in phenotypes to verify if different sets of parameters could characterize different subtypes of parkinsonians; other balance tests should be considered to assess balance instability. These will be the next steps of our activity; the current findings encourage us to continue along this line of research to achieve a comprehensive system for the automatic and reliable assessment of PD status, suitable for the home monitoring of disease progression.
