Toward interpretability of machine learning methods for the classification of patients with major depressive disorder based on functional network measures.

We address the interpretability of the machine learning algorithm in the context of the relevant problem of discriminating between patients with major depressive disorder (MDD) and healthy controls using functional networks derived from resting-state functional magnetic resonance imaging data. We applied linear discriminant analysis (LDA) to the data from 35 MDD patients and 50 healthy controls to discriminate between the two groups utilizing functional networks' global measures as the features. We proposed the combined approach for feature selection based on statistical methods and the wrapper-type algorithm. This approach revealed that the groups are indistinguishable in the univariate feature space but become distinguishable in a three-dimensional feature space formed by the identified most important features: mean node strength, clustering coefficient, and the number of edges. LDA achieves the highest accuracy when considering the network with all connections or only the strongest ones. Our approach allowed us to analyze the separability of classes in the multidimensional feature space, which is critical for interpreting the results of machine learning models. We demonstrated that the parametric planes of the control and MDD groups rotate in the feature space with increasing the thresholding parameter and that their intersection increases with approaching the threshold of 0.45, for which classification accuracy is minimal. Overall, the combined approach for feature selection provides an effective and interpretable scenario for discriminating between MDD patients and healthy controls using measures of functional connectivity networks. This approach can be applied to other machine learning tasks to achieve high accuracy while ensuring the interpretability of the results.