Decision-based data fusion of complementary features for the early diagnosis of Alzheimer's disease

As the average life expectancy increases, particularly in developing countries, the prevalence of Alzheimer's disease (AD), which is the most common form of dementia worldwide, has increased dramatically. As there is no cure to stop or reverse the effects of AD, the early diagnosis and detection is of utmost concern. Recent pharmacological advances have shown the ability to slow the progression of AD; however, the efficacy of these treatments is dependent on the ability to detect the disease at the earliest stage possible. Many patients are limited to small community clinics, by geographic and/or financial constraints. Making diagnosis possible at these clinics through an accurate, inexpensive, and noninvasive tool is of great interest. Many tools have been shown to be effective at the early diagnosis of AD. Three in particular are focused upon in this study: event-related potentials (ERPs) in electroencephalogram (EEG) recordings, magnetic resonance imaging (MRI), as well as positron emission tomography (PET). These biomarkers have been shown to contain diagnostically useful information regarding the development of AD in an individual. The combination of these biomarkers, if they provide complementary information, can boost overall diagnostic accuracy of an automated system. EEG data acquired from an auditory oddball paradigm, along with volumetric T2 weighted MRI data and PET imagery representative of metabolic glucose activity in the brain was collected from a cohort of 447 patients, along with other biomarkers and metrics relating to neurodegenerative disease. This study in particular focuses on AD versus control diagnostic ability from the cohort, in addition to AD severity analysis. An assortment of feature extraction methods were employed to extract diagnostically relevant information from raw data. EEG signals were decomposed into frequency bands of interest hrough the discrete wavelet transform (DWT). MRI images were reprocessed to provide volumetric representations of specific regions of interest in the cranium. The PET imagery was segmented into regions of interest representing glucose metabolic rates within the brain. Multi-layer perceptron neural networks were used as the base classifier for the augmented stacked generalization algorithm, creating three overall biomarker experts for AD diagnosis. The features extracted from each biomarker were used to train classifiers on various subsets of the cohort data; the decisions from these classifiers were then combined to achieve decision-based data fusion. This study found that EEG, MRI and PET data each hold complementary information for the diagnosis of AD. The use of all three in tandem provides greater diagnostic accuracy than using any single biomarker alone. The highest accuracy obtained through the EEG expert was 86.1 ±3.2%, with MRI and PET reaching 91.1 +3.2% and 91.2 ±3.9%, respectively. The maximum diagnostic accuracy of these systems averaged 95.0 ±3.1% when all three biomarkers were combined through the decision fusion algorithm described in this study. The severity analysis for AD showed similar results, with combination performance exceeding that of any biomarker expert alone.

Name thesis:degree_name

M.S. Engineering

Level thesis:degree_level

Thesis

Discipline thesis:degree_discipline

Electrical and Computer Engineering

Year dc:date.available

2009

Author dc:creator

• Ahiskali, Metin

Contributors dc:contributor

• Polikar, Robi

• Biochemical markers; Alzheimer's disease; Diagnostic imaging
• Electrical and Computer Engineering

Repository record dc:identifier

https://rdw.rowan.edu/etd/21

OAI identifier oai:identifier

oai:rdw.rowan.edu:etd-1020