The Solution to the Problem of Classifying High-Dimension fMRI Data Based on the Spark Platform

Abstract

This paper compares approaches to solving the classification problem based on fMRI data of the original dimension using the big data platform Spark. The original data is 4D fMRI time series with time resolution (TR) = 0.5 s for one sample recording. Participants have to solve 6 tasks, requiring activating various types of thinking, during 30 min session. A large number of subjects and a short time resolution generated the dataset with more than 86 000 samples, which allowed applying machine learning methods to solve this problem, instead of classical statistical maps. The random forest model was used to solve the binary classification problem. The paper analyzes model performance dependence upon time during the problem solving sessions. Evidence has been obtained that there is some limited time required for solving the same type of problems, and if more time is spent, this is due to the fact that the brain does not instantly get involved in the work on the proposed task, but it is still staying at resting state for some time.