MAMF-GCN model for anxious and non-anxious depression classification and neuroimaging marker recognition

Depression is a prevalent psychological disorder, and distinguishing anxious from non-anxious depression and identifying neuroimaging markers are challenges.

In this study, we used public data from the REST-meta-MDD consortium, selecting 158 patients with anxious depression and 108 with non-anxious depression. We computed multiple neuroimaging metrics, including fractional amplitude of low-frequency fluctuations (fALFF), degree centrality (DC), regional homogeneity (ReHo), voxel-mirrored homotopic connectivity (VMHC), and functional connectivity (FC). We proposed the multi-scale adaptive multi-channel fusion deep graph convolutional network (MAMF-GCN) for anxious and non-anxious depression classification. This model incorporated both neuroimaging-derived features and relevant non-imaging clinical information, including site, age, gender, and Hamilton Depression Rating Scale (HAMD) scores. Furthermore, we applied the gradient backpropagation module to screen potential neuroimaging markers for depression subtyping.

Our approach achieved outstanding classification results, with accuracy (ACC) of 98.12 %, sensitivity (SEN) of 99.41 %, specificity (SPE) of 97.46 %, F1-score of 98.38 %, and area under the curve (AUC) of 0.9867. Compared to state-of-the-art methods, the proposed MAMF-GCN model demonstrated superior performance. Ablation experiments confirmed that the MAMF-GCN model effectively utilizes non-imaging clinical information to significantly improve classification performance. Utilizing the gradient backpropagation module, we successfully pinpointed the top ten most discriminative features. Furthermore, we identified the cerebellar network as playing a critical role in differentiating anxious from non-anxious depression, with functional connectivity in the visual, auditory, and motor control networks serving as potential markers.

This method highly effective in classifying anxious and non-anxious depression and in identifying potential neuroimaging markers, thus holding significant clinical application value.