Discovering Knowledge of ASD from CCC-2: Ensemble Learning Approach for Analysis of ASD

In this paper, we constructed an ASD classifier by random forest with responses of CCC-2 and the diagnosis results obtained from ADOS. Further the importance of features in CCC-2 for the classification of ASD was analyzed. The hyperparameters of the random forest were adjusted on the training dataset with the cross-validation, and the generalization performance was evaluated on the test dataset. Since the sample size was not so large, we validated the effect of random shuffling for the classification performance with additional 4 shuffle pattern. The all constructed classifiers not only had a highly classification performance, but also the result was stable with respect to random shuffling. It is also remarkable result that two items, which related to pragmatic impairments, were consistently determined to be the first, second important feature respectively. The items that reflect these pragmatic impairments were emphasized over the I and J domains in CCC-2, which reflect the main behavioral characteristics of ASD. It shed light on new aspects of ASD assessment for children.