TITLE:
Predictive Models for Functional MRI Data
AUTHORS:
Guenadie Nibbs, Peter Bajorski
KEYWORDS:
Functional Magnetic Resonance Imaging, Regression, Logistic Regression, Linear Discriminant Analysis, Random Forest
JOURNAL NAME:
Open Journal of Statistics,
Vol.10 No.1,
January
9,
2020
ABSTRACT: In this
study, we analyze brain activity data describing functional magnetic resonance
imaging (MRI) imaging of 820 subjects with each subject being scanned at 4
different times. This multiple scanning gives us an opportunity to observe the
consistency of imaging characteristics within the subjects as compared to the
variability across the subjects. The most consistent characteristics are then
used for the purpose of predicting subjects’ traits. We concentrate on four
predictive methods (Regression, Logistic Regression, Linear Discriminant
Analysis and Random Forest) in order to
predict subjects’ traits such as gender and age based on the brain activities observed
between brain regions. Those predictions are done based on the adjusted
communication activity among the brain regions, as assessed from 4 scans of
each subject. Due to a large number of such communications among the 116 brain
regions, we performed a preliminary selection of the most promising pairs of
brain regions. Logistic Regression performed best in classifying the subject gender based on communication activity
among the brain regions. The accuracy rate was 85.6 percent for an AIC
step-wise selected Logistic Regression model. On the other hand, the Logistic
Regression model maintaining the entire set of ranked predictor was capable of getting an 87.7 percent accuracy rate. It is
interesting to point out that the model with the AIC selected features was
better classifying males, whereas the complete ranked model was better classifying
females. The Random Forest technique performed best for prediction of age
(grouped within five categories as provided by the original data) with 48.8 percent accuracy rate. Any set of
predictors between 200 and 1600 was presenting similar rates of accuracy.