Deep graph neural network-based prediction of acute suicidal ideation in young adults

Figures

• 

  Figure 1. Overall architecture of a prediction model based on a graph isomorphism network (GIN): (a) Overall workflow for constructing the graph-structured dataset. A single subject corresponds to a single graph. (b) Graph constructed on multi-dimensional self-report questionnaires and clinico-demographic information, which were used as input node features and fed into five graph convolutional layers combined with MLP layers. In each layer, the graph representation is obtained after graph pooling, concatenated into a latent feature vector, and fed into a classifier using a fully connected layer to output a sigmoid prediction of MaDEs or acute SI score (0–1). (c) Three different GIN models with different subsampling strategies (i.e., GIN-SMOTE, GIN-u1, and GIN-u2) were ensembled to obtain the final model to overcome class imbalance. (d) Sparse correlation network for the edge matrix of the graph: pairwise correlation coefficients were obtained between the categorical variables of each node, representing each questionnaire item and subject information feature and were used as an edge matrix to construct a graph. All subjects share the same edge matrix to construct a graph. Note that PHQ-9…

• 

  Figure 2. Attention plots: (a) Raw averaged attention plots, (b) Attention plot comparing the questionnaire items, (c) Attention plot comparing the binary items, (d) Attention plot for the mixed Gaussian model-based graphical network (generated with the “mgm” R package for visualization) for the questionnaire items on the test set (n = 14,238).

• 

  Figure 3. Receiver operating characteristic (ROC) curves of models using (a) u1, (b) u2, (c) SMOTE, and (d) ensemble of u1, u2, and SMOTE for ablation study. There were significant but small differences between the diagnostic performance of the model with and without Item 9 of PHQ-9 except for u2. Each p-value was calculated from DeLong’s test comparing two ROC curves.

• 

  Figure 4. Scatter plot showing the normalized raw prediction score of the model and KSSI total score (n = 792 of 13,408; the part of the test set with KSSI scores). Spearman’s correlation coefficient between the prediction score of the model and the KSSI total score was \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\rho }_{\mathrm{pred}}$$\end{document}ρpred = 0.599 (p < 0.0001).