Abstract
With the release of the multi-site Autism Brain Imaging Data Exchange, many researchers have applied machine learning methods to distinguish between healthy subjects and autistic individuals by using features extracted from resting state functional MRI data. An important part of applying machine learning to this problem is how to extract these features. Specifically, whether to include anti-correlations between brain regions as relevant features and how best to define these features. For the second question, the graph theoretical properties of the brain network may provide a reasonable answer. In this study, we investigated the first issue by comparing three different approaches. These included using the unaltered correlation matrix, the absolute value of the correlation matrix or the inverted correlation matrix (negative of the original) as the starting point for extracting relevant features using graph theory. We then trained a multi-layer perceptron in a leave-one-site out manner in which the data from a single site was left out as testing data and the model was trained on the data from the other sites. Our results show that on average, using graph features extracted from the anti-correlation-based matrix led to the highest accuracy and AUC scores. We also show that adding the PCA transformation of the original correlation matrix to the feature space leads to an increase in accuracy. This suggests that anti-correlations should not simply be discarded as they may include useful information that would aid the classification task.