Prediction of neurodegenerative diseases from functional brain imaging data

Neurodegenerative diseases are a challenge, especially in the
developed society where life expectancy is high. Since these diseases
progress slowly, they are not easy to diagnose at an early
stage. Moreover, they portray similar disease features, which makes
them hard to differentiate. In this thesis, the objective was to
devise techniques to extract biomarkers from brain data for the
prediction and classification of neurodegenerative diseases, in
particular parkinsonian syndromes. We used principal component
analysis in combination with the scaled subprofile model to extract
features from the brain data to classify these disorders. Thereafter,
the features were provided to several classifiers, i.e., decision
trees, generalized matrix learning vector quantization, and support
vector machine to classify the parkinsonian syndromes. A validation
of the classifiers was performed.

The decision tree method was compared to the stepwise regression
method which aims at linearly combining a few good principal
components. The stepwise regression method performed better than the
decision tree method in the classification of the parkinsonian
syndromes. Combining the two methods is feasible. The decision trees
helped us to visualize the classification results, hence providing an
insight into the distribution of features. Both generalized matrix
learning vector quantization and support vector machine are better
than the decision tree method in the classification of early-stage
parkinsonian syndromes.

All the classification methods used in this thesis performed well with
later disease stage data. We conclude that generalized matrix learning
vector quantization and decision tree methods can be recommended for
further research on neurodegenerative disease classification and
prediction.