Digital Theses Archive


Tesi etd-08032018-172423

Type of thesis
Study and development of novel diagnostic and therapeutic approaches for Parkinson's disease based on ICT and machine learning
Scientific disciplinary sector
INGEGNERIA - Biorobotics
relatore Dott. CAVALLO, FILIPPO
  • Early Diagnosis
  • Machine Learning
  • Motion Analysis
  • Parkinson's Disease
  • Signal Processing
  • Wearable Inertial Sensors.
Exam session start date
The rapidly ageing of the population is causing an increase of people affected by neurodegenerative diseases. Among these, Parkinson’s Disease (PD) is a complex disorder caused by critical loss of dopamine in the forebrain, which afflicts approximately 6.5 millions of people worldwide. PD is highly disabling for patients because of its cardinal motor symptoms (tremor, muscular rigidity, bradykinesia, postural instability) and non-motor manifestations (e.g., hyposmia, sleep disorders, autonomic dysfunctions). Since current diagnostic methods and disease progression monitoring approaches are suboptimal for PD management, the need to redefine the research criteria for the diagnosis of this complex disease has been recently recognized.<br>On the other hand, the growing interest of high-tech companies in healthcare leads to study Information and Communication Technologies (ICTs), as well as Machine Learning techniques, to develop integrated solutions, demonstrating their feasibility, effectiveness, acceptability, and sustainability for clinical applications.<br>In this context, this dissertation aims to propose a wearable system, based on inertial measurement unit (IMU) technologies and custom-made algorithms for data processing, to support clinicians in the diagnosis and monitoring of PD. Wearable sensors, indeed, are more and more used for healthcare applications thanks to their good trade-off between unobtrusiveness and accuracy of measurements, and optimized data processing are needed to manage the great amount of information they can acquire.<br>The idea is to provide the neurologist with a non-invasive easy to use wireless system, composed by SensHand and SensFoot devices and related processing algorithms, able to objectively and accurately measure the subjects motor performance. Currently, in fact, PD diagnosis is mainly based on the visual assessment of subjects motion during the performance of standardized clinical tasks, thus diagnosis is affected by intra and inter-rater variability. In particular, such a system would act as support for clinicians to address three main clinical challenges: i) the identification of a subtle deflection in motor capabilities that allows detecting the disease in the prodromal stage; ii) the accurate quantification of the disease motor assessment in PD patients to objective PD diagnosis and pathology stage; and iii) the objective evaluation of the response to pharmacological therapies.<br>A well-structured experimental protocol was defined, in synergy with the clinical staff, to enable full evaluations of motor performance. The protocol was initially based on tasks by MDS-UPDRS III, since the compliance of the system to clinical evaluations should be firstly demonstrated. Then, also other similar exercises were added for their potential ability to improve PD symptoms assessment. Several experimental sessions were set up, and a sizeable amount of motor data about PD patients, healthy subjects of control (HC), and subjects with idiopathic hyposmia (IH) was collected.<br>Since the accuracy of measurement of the system is one of the most limiting barriers to overthrow to enable the effective use of novel technology in clinical practice, a preliminary validation of the system was carried out. Extracted features from the wearable system were compared to those obtained by an optoelectronic system, which is the &#34;gold standard&#34;. Very low discrepancies were found between the two measurement systems, which resulted in highly correlated measures. Then, from a clinical point of view, it was studied whether the extracted features were repeatable and significant to discriminate between a PD group and a control group. Even if these results should be confirmed on a larger dataset, this is a first step toward the clinical validation of the system to accurately identify PD signs.<br>Then, the wearable system, meant as inertial sensors and processing algorithms,<br>was tested on different applications, for addressing the stated clinical challenges.<br>Importantly, it was used for a case study within a two-step approach, following a validate olfactory test, as a possible way for early detecting PD. The system allowed to identify significant variations of motion comparing the motor performance of an IH subject at baseline and after one year. Then the system enabled the objective monitoring of the motor performance in that IH subject when a dopaminergic challenge test was administered him (Clinical Challenge 3). The positive response to the pharmacological test induced to suspect the IH subject for possible PD diagnosis that was confirmed by [123I]-FP/CIT SPECT. Therefore, the wearable system resulted suitable for identifying subtle changes of motion that can lead to hypothesize preclinical PD diagnosis (Clinical Challenge 1).<br>Furthermore, machine learning techniques were applied to investigate whether the<br>acquisitions made with the inertial system during the performance of MDS-UPDRS III tasks, allowed: i) the extraction of a feature array of motor parameters able to identify significant differences between three groups of people (i.e., HC, IH, and PD) by using supervised classifiers (Clinical Challenges 1-2); ii) the objective diagnosis of PD according to the pathology level by implementing unsupervised learning (Clinical Challenge 2). Good results were achieved in all applications, and putting information from all four limbs together outperformed the analyses conducted on lower and upper limbs separately. This result highlights the importance to have a comprehensive protocol which includes tasks for a full assessment of subjects’ motion.<br>Despite improvements could be achieved with further studies (enlarged dataset,<br>real-time processing, follow-up analysis), this dissertation is a first step toward using the proposed system in clinical practice with the aim to actually support clinicians not only in objective quantification of PD motor signs (Clinical Challenge 2), but also in identification of people at risk for developing PD, allowing to follow-up that subjects over the time to reveal the pathology in a prodromal stage (Clinical Challenge 1).<br>