Graduation Year


Document Type




Degree Name

Doctor of Philosophy (Ph.D.)


Electrical Engineering

Degree Granting Department

Electrical Engineering

Major Professor

Ravi Sankar, Ph.D.

Committee Member

Stephen Saddow, Ph.D.

Committee Member

Ismail Uysal, Ph.D.

Committee Member

Marvin Andujar, Ph.D.

Committee Member

Kandethody Ramachandran, Ph.D.


Parkinson’s disease, Event related potential, Brain computer interface, Brain source localization, Higher order statistics, Bispectrum, Bicepstrum, Quadratic phase coupling


The field of signal processing has many applications, one of which is in the field of biomedical engineering where it has improved the performance of biomedical devices and the accuracy of medical diagnosis. One of the areas that have benefited from this field is the diagnosis of Parkinson’s disease (PD). This disease is one of the most common neurodegenerative diseases of the central nervous system. Nearly one million Americans suffer from PD, and this number goes up to over ten million people worldwide. The main symptoms of PD include bradykinesia, rest tremor, rigidity, and impaired balance. There are many types of biomedical data that have been used in the diagnosis process of PD, however, most of the applied biomedical signals rely on the presence of motor symptoms which means in most cases, by the time that the patients are diagnosed, they are likely to have lost the majority of the dopaminergic neurons in their brains.

One of the signals that have been used for the diagnosis of PD is electroencephalography (EEG). The neurons in the brain communicate with each other through electrical potentials that appear at the synapses. EEG is a noninvasive method that collects the small voltages that appear on the scalp caused by large clusters of neurons using multiple electrodes; therefore, EEG recordings are multi-channel signals where each channel is corresponding to a specific region of the brain. There are two main types of EEG signals, background EEG, and event-related potential (ERP). Background EEG is the signal collected during the rest state and contains the regular activity of the brain when it is not provoked, whereas ERP is the changes in the background EEG, resulting from a specific stimulus. While background EEG signals are better suited for diagnosis purposes, they are highly nonlinear, non-stationary and non-Gaussian signals; hence, to extract relevant information from them, advanced methods of signal processing are required. The background EEG is a random signal which indicates that regular features such as time locked features or peaks of the signal do not carry much information. For random signals, statistics is usually the most appropriate method for analysis.

The word statistic is usually used to refer to first and second order statistics. Higher order statistics (HOS) is defined as a more general term that covers higher order of statistical features. the field of HOS analysis is usually employed for highly complex signals, where the first and second order statistics failed to adequately define the system. Due to the highly random nature of background EEG signals, HOS has been employed by many researchers for more detailed analysis. In this study, a range of HOS features have been used to improve the diagnosis performance of PD patients from healthy control (HC) and classification of stages of PD after a positive diagnosis. A detailed analysis of the features was performed to find the best combination for this application and a number of new HOS features were developed to improve the performance of the model. Concurrently, based on previous research a spectral analysis of the data was performed to investigate the effect of PD on brain rhythms where HOS features were extracted from multiple rhythms and used separately in the diagnosis process.

The classification stage was performed by a range of conventional, ensemble and deep learning algorithms while employing the leave-one-trial-out (LOTO), leave-one-subject-out (LOSO) cross validation (CV) methods. To preserve the balance of the data, a new CV approach, leave-two-subjects-out (LOTO) was also employed (one from each class). For diagnosis of PD, a comparison between the features extracted from different brain rhythms and different classifiers was performed. The result was then compared to several deep learning methods and other state-of-the-art approaches. The performance of the PD stage classification was also compared to other studies in this field. Together these two methods create a unified hierarchy model to diagnose and identify the stages of PD.