Date Approved

12-31-2004

Embargo Period

4-25-2016

Document Type

Thesis

Degree Name

M.S. in Electrical Engineering

Department

Electrical & Computer Engineering

College

Henry M. Rowan College of Engineering

Advisor

Polikar, Robi

Subject(s)

Alzheimer's disease; Electroencephalography; Wavelets (Mathematics)

Disciplines

Electrical and Computer Engineering

Abstract

Alzheimer's disease is a neurological disorder characterized by nerve degeneration and neuronal death. The diagnosis of Alzheimer's disease at an early stage is a major concern due to the lack of a standard and effective diagnosis procedure available to community healthcare providers, as well as the increasing numbers of the elderly population affected. Clinical evaluation, the standard AD diagnostic procedure conducted at major university hospitals and research clinics, achieves, on average, a positive predictive value of 93%, with a sensitivity of 83%, specificity of 55% and an overall accuracy of 75%. An effective and objective tool for early diagnosis of the disease is important to have a meaningful impact on healthcare. Such a procedure must be inexpensive, non-invasive and available to community physicians, who often provide the first line of intervention, particularly at the early stages of the disease.

Studies performed using wavelets or other signal processing methods to analyze EEG signals in an attempt to find a non-invasive biomarker for Alzheimer's disease have had varying degrees of success. Two types of wavelets have commonly been used for analyzing the event-related potentials of EEG signals: Daubechies 4 wavelets and Quadratic B-spline wavelets. Analysis was performed using these two types of wavelets and the coefficients obtained from this analysis were then used with several algorithms for automated classification. Cross validation was performed with a multilayer perceptron (MLP) neural network and an ensemble of MLP networks for classification of the coefficients. The classification algorithms are compared as well as the two types of wavelets. Noting that different information may be available from two different types of wavelets, a data fusion method with an ensemble of classifiers was implemented to combine relevant information and possibly boost performance of the algorithm.

The use of an automated algorithm is a feasible approach for diagnosing AD in a community health clinic. The accuracy of each method used in this study is similar to that of a clinical evaluation and as an automated algorithm provides a diagnostic tool for the detection of AD that can be made available to community physicians.

The results for data fusion indicate that the wavelets are not likely extracting complementary information from the signals since the combination of these two wavelet feature sets does not consistently produce a more informed decision than either set individually. The use of data fusion for combining features is however, a feasible approach to this classification problem if the chosen feature sets provide complementary information.

Download

Share

COinS