Comparative Study of Dimensionality Reduction Methods Using Reliable Features for Multiple Datasets Obtained by rs-fMRI in ADHD Prediction

摘要

ADHD is the most commonly diagnosed psychiatric disorder in children and, although its diagnosis is done in a subjective way, it can be characterized by abnormality work of specific brain regions. Datasets obtained by rs-fMRI cooperate to the large amount of brain information, but they lead to the curse-of-dimensionality problem. This paper aims to compare dimensionality reduction methods belonging to feature selection task using reliable features for multiple datasets obtained by rs-fMRI in ADHD prediction. Experiments showed that features evaluated in multiple datasets were able to improve the correct labeling rate, including the 87% obtained by MRMD that overcomes the higher accuracy in rs-fMRI ADHD prediction. They also eliminated the curse-of-dimensionality problem and identified relevant brain regions related to this disorder.