Quantitative assessment of lower limb bradykinesia in Parkinson's disease patients using smartphone sensors.

In a perspective of evidence-based medicine, data from measurement systems, processed by data mining techniques, can provide quantitative measures that can make more objective the available clinical information and be an important aid for clinical decision making. This is the leitmotif of this project. In this Master Thesis work the focus is on motor symptoms of Parkinson's disease, in particular on bradykinesia (slowness of movement). The motivation is found in the need to have quantitative methods for the motor assessment of patients, in order to be able to monitor the daily fluctuations of the motor symptoms (alternate of phase ON and phase OFF). In fact, the most used clinical evaluation scale, the MDS-UPDRS, in which the clinician is asked to assign a score between 0 and 4 according to the severity of the considered symptom, does not meet requirements of objectivity and repeatability. Furthermore, the fact that this kind of evaluation is performed only within follow-up sessions make impossible for the neurologist to observe the variations of the pharmacological/surgical treatment that occurs in Parkinson's disease patients. The main techniques for the assessment of motor functions in Parkinson's disease, based on inertial sensors located in wearable devices, were analysed in literature. The purpose of this study is to investigate the use of a common smartphone to collect data during the execution of a motor task and the best machine learning method for the quatitative assessment of bradykinesia. Ninety-three Parkinson's disease patients and ten healthy people partecipated in this study. The data used in the project was gathered from a smartphone positioned on the thigh while the patient conducted a prescribed movement activity: the Leg Agility (task 3.8 in the MDS-UPDRS). Some machine learning models were trained and tested for the supervised classification based on the UPDRS scale. A set of 16 features were extracted from the data and both multiclass classification and binary classification were performed. The better overall correct classification percentages are reached by SVM model with 60.9 percent and Neural Network model with 76.1 percent. The most accurate classification was achived for observations with true class 1 and 3. The kNN, Linear Discriminant Analysis and Decision Tree show their inability to recognize the cases of UPDRS 4.