A Multimodal Neuroimaging Approach for Classification and Prediction of Alzheimer's Disease Using Machine Learning

Authors

Parisa ForouzannezhadFollow

Document Type

Dissertation

Degree

Doctor of Philosophy (PhD)

Major/Program

Electrical and Computer Engineering

First Advisor's Name

Malek Adjouadi

First Advisor's Committee Title

Committee chair

Second Advisor's Name

Armando Barreto

Second Advisor's Committee Title

Committee Member

Third Advisor's Name

Jean Andrian

Third Advisor's Committee Title

Committee Member

Fourth Advisor's Name

Mercedes Cabrerizo

Fourth Advisor's Committee Title

Committee Member

Fifth Advisor's Name

Naphtali Rishe

Fifth Advisor's Committee Title

Committee Member

Keywords

electrical and computer engineering

Date of Defense

11-13-2020

Abstract

Alzheimer’s disease (AD) is one of the most common neurodegenerative disorders among the elderly population. It is progressive, irreversible in nature, and is considered the main cause of dementia. AD has become a world health problem affecting developed and developing nations alike, with the number of diagnosed AD patients increasing rather dramatically as both the life span of humans and the earth’s population continue to increase. Therefore, AD diagnosis in its earliest manifestations, preferably at the presymptomatic stage is critical for the timely planning of treatment and therapeutic interventions. We introduce new machine learning algorithms to detect and predict Alzheimer’s disease in the early phase to include the presymptomatic stage where no manifestation of cognitive decline is yet apparent. An investigation is carried out in search of optimal feature selection methods on different machine learning platforms with the intent to address the challenging classification and regression analysis. This research endeavor introduces three machine learning platforms that are based on (1) deep neural network, (2) support vector machine (SVM), and (3) Gaussian-based model classifiers all optimized in order to delineate the different stages of the disease as well as a regression framework to predict future cognitive scores as means to gauge disease progression, which could play an important role in pre-and post-treatment evaluations. The input data to these machine learning architectures included magnetic resonance imaging (MRI), positron emission tomography (PET), the metabolic fluorodeoxyglucose (FDG)-PET, cognitive scores, cerebrospinal fluid (CSF), and the apolipoprotein E4 (APOE4) gene. An investigation is carried out on the transition phases of AD through regression analysis by predicting cognitive tests including Alzheimer’s disease assessment scale cognitive subscale (ADAS-Cog), Mini-mental state examination (MMSE), and Rey’s auditory verbal learning (RAVLT) that have been designed and used as important criteria to evaluate cognitive status of AD patients. We formulated the prediction of disease progression as a multimodal multitask regression problem across six time points in a 4-year longitudinal study. Major findings of this work reveal that for binary classification, the highest accuracy of 84% for delineating the challenging group of early mild cognitively impaired individuals (EMCI) from the cognitively normal (CN) group is obtained. With multiclass classification using deep neural network methodology, especially when early and late MCI (EMCI and LMCI) groups are included, the accuracy does not exceed 70%, which clearly explains the many nuances in the transition phases of the disease, especially in the early stages. Moreover, the episodic tests like RVALT as used in this study were shown to be effective for selecting the at-risk groups. MRI morphometry was found to be the most sensitive biomarker to predict disease conversion and observed that parietal and prefrontal cortices are also associated with episodic memory in addition to the temporal lobe. Although adding the modalities of FDG-PET, CSF, and APOE allele gene improved the prediction error significantly at 4 time points, multimodal neuroimaging does not statistically enhance the prediction performance at some time points due to the inherent challenge of missing data. It is clear that for longitudinal studies of such duration (4-year), beyond the variability and interrelatedness of features, the missing data challenge remains the most difficult to overcome.

Identifier

FIDC009203

Recommended Citation

Forouzannezhad, Parisa, "A Multimodal Neuroimaging Approach for Classification and Prediction of Alzheimer's Disease Using Machine Learning" (2020). FIU Electronic Theses and Dissertations. 4584.
https://digitalcommons.fiu.edu/etd/4584