Сборник статей конференции "Информационные технологии и системы" (ИТиС'16)
In this work, graph spectra of the normalized graph Laplacians of brain networks (connectomes) are used for solving the task of classifying autism spectrum disorder against typical development. We find the most informative group of eigenvalues by introducing a window and sliding it through all possible positions. We next assume that these values are sampled from a Dirichlet distribution and build a linear model with a single feature that is based on estimation of a Dirichlet parameter. The proposed classifier outperforms the baseline in terms of both mean ROC AUC value (0.74) and stability of ROC AUC values to the variations in the data. Classifiers that implemented a similar approach but used geometric distances instead of statistical methods showed worse performance. This implies that the Dirichlet distribution might be a useful tool for the analysis of normalized Laplacian spectra when solving tasks of classifying brain networks.
We solve a task of classifying autism spectrum disorder versus normal controls based on structural brain networks (connectomes). We compare different approaches to machine learning in a nspecial conditions when each subject is represented by a set of connectomes obtained after applying different weightings and normalizations to the original dataset. We consider two algorithms of two-level classifications: stacking and blending of the models trained on the different types of the data. We also build a discriminative fusion classifier which is a logistic regression on the weighted combination of the varios types of connectomes. The best classification quality (ROC AUC of 0.8) is obtained for blending which is a weighted combination of the logistic regression models; this model performes better than the first-level individual models.