Using Convolutional Neural Networks and Connectome-Based Predictive Modeling to Predict ADHD Diagnoses
Determining how to best predict Attention Deficit Hyperactivity Disorder (ADHD) can contribute to our understanding of its neurological basis. Previous work suggests that both connectome-based predictive models (CPMs) and connectome-convolutional neural networks (CCNNs) can be effective tools to predict ADHD from functional Magnetic Resonance Imaging (fMRI) scans. I used both types of models and 10-fold crossvalidation to predict the ADHD Rating Scale IV scores of 229 subjects based on their functional connectivity patterns measured with resting-state fMRI data. I hypothesized that CCNNs would have stronger predictive power than CPMs, because of the former’s ability to handle both complex patterns and sparse data. I found, however, that although both methods predicted ADHD symptoms well, CPMs outperformed CCNNs. I also discovered that models performed better when functional connectivity was measured with Pearson correlation rather than with dynamic time warping distance. Moreover, there are high average correlations between the predicted scores of models with varying input and method types. These high correlations suggest that both input types encompass similar information and that both method types pick up on certain common trends in the data. Overall, these results suggest that CPMs deal with the challenges of fMRI data better than CCNNs do.
Gigi Stark