Motor fluctuations are a major focus of clinical managements in patients with mid-stage and advance Parkinson's disease (PD). In this thesis, an automated algorithm is developed to identify those fluctuations (i.e., medication OFF and ON) using wearable sensors while PD patients are engaging in a variety of daily living activities. Four different methods are proposed which are supervised learning using Support Vector machine (SVM) with fuzzy classification, semi-supervised learning using k-means or using Self-organizing Tree Map Algorithm (SOTM) with fuzzy classification, and supervised classification using Long short-term memory (LSTM) as a deep learning method.

A set of temporal and spectral features are extracted from the ambulatory signals of triaxial gyroscope sensors. After performing dimensionality reduction, the features are introduced to SVM or clustering methods using k-means or SOTM. Signals of the gyroscope sensors are passed directly to LSTM network. The developed methods were evaluated on two datasets that included recordings of 19 PD patients. Two scenarios were considered: general training/classification and patient-specific where the former trains and tests the algorithm using subject-based leave-one-out cross-validation, and the latter trains and tests the algorithm for each patient individually. In addition, for patient-specific scenarios, the number and placement of sensors is selected for each patient and this selection is based on the average change in UPDRS score between ON and OFF medication states and the presence of tremor for that patient.

Overall, patient-specific algorithm resulted in a higher classification performance when it based on SVM with fuzzy classification (i.e., 80%, 82% and 78% for accuracy, sensitivity, and specificity of the OFF state, respectively). This algorithm was able to classify the medication states with high confidence (i.e., accuracy 94.86%, sensitivity 91.94% and specificity 96.83%) for the group of patients with a change of more than 15 in their UPDRS score between the OFF and ON medication states. This results are promising and thus this algorithm can be potentially used in routine clinical practice to improve the quality of this group of PD patients.

In addition to these results, when only one sensor mounted on the ankle was used in the general training scenario, the algorithm based on LSTM performed better than SVM with 74.91%, 69.42%, and 80.55% for accuracy, sensitivity, and specificity, respectively. The promising results of LSTM show the potential outcome of developing deep learning methods in this field.

Library of Congress Subject Headings

Parkinson's disease--Patients--Rehabilitation; Drugs--Measurement

Publication Date


Document Type


Student Type


Degree Name

Computer Engineering (MS)

Department, Program, or Center

Computer Engineering (KGCOE)


Behnaz Ghoraani

Advisor/Committee Member

Andreas Savakis

Advisor/Committee Member

Andres Kwasinski


Physical copy available from RIT's Wallace Library at RC382 .H77 2017


RIT – Main Campus

Plan Codes