Comparing Functional and Structural Predictors of Cognition via Machine Learning
G. Andrew James1, Ming-Hua Chung1, Keith A. Bush1, Clinton D. Kilts1; 1University of Arkansas for Medical Sciences
Neuroscience has long sought to identify the structural and functional correlates of normative cognition, but few datasets have allowed robust comparison of these modalities’ predictive ability. The recent growth of large neuroimaging initiatives allows unprecedented opportunities to evaluate brain-cognition relationships. We selected N=982 participants (53% female; ages 21-35) from the WU-Minn Human Connectome Project with complete structural MRI (sMRI), diffusion tensor imaging (DTI), and resting-state functional MRI (rs-fMRI) neuroimaging datasets. The machine learning tool LASSO (least absolute shrinkage and selection operator) was implemented to linearly regress features from each neuroimaging modality (X: sMRI, DTI, rs-fMRI) to seven cognitive measures (Y) derived from the NIH Toolbox. Participants were randomly split into training and testing datasets (N=788 and N=194), stratified by gender and age. For each modality (X) and cognition (Y), LASSO was implemented with 10-fold cross-validation of the training dataset to identify the beta-weights of predictors which best predicted cognition (i.e. the model with minimum MSE). We compared percent variance explained (R2) by each modality-cognition pair for both training and testing datasets. For the training dataset, rs-fMRI explained greater variance in cognition (R2 0.05-0.36) than sMRI (R2 0.00-0.10) or DTI (R2 0.00-0.08; F(2,18)=8.38, p<0.003). rs-fMRI also explained greater variance in the testing dataset (R2 0.00-0.17) than sMRI or DTI (F(2,18)=3.80, p<0.05). However, rs-fMRI explained significantly less variance in testing dataset than training (paired t(6)=4.5, p<0.01), indicating poor generalization possibly due to overfitting. Future work will explore LASSO optimization approaches to improve generalization of findings.
Topic Area: METHODS: Neuroimaging