Nonsuicidal self-injury prediction with pain-processing neural circuits using interpretable graph neural network

Abstract

Background: Nonsuicidal self-injury (NSSI) involves the intentional destruction of one's own body tissues without suicidal intent. Prior research has shown that individuals with NSSI exhibit abnormal pain perception; however, the pain-processing neural circuits underlying NSSI remain poorly understood. This study leverages graph neural networks to predict NSSI risk and examine the learned connectivity of neural underpinnings using multimodal data.

Methods: Resting-state functional MRI and diffusion tensor imaging were collected from 50 patients with NSSI, 79 healthy controls (HC), and 44 patients with mental disorder who did not engage in NSSI as disease controls (DC). We constructed pain-related brain networks for each participant. An interpretable graph attention networks (GAT) model was developed, considering demographic factors, to predict NSSI risk and highlight NSSI-specific connectivity using learned attention matrices.

Results: The proposed GAT model based on imaging data achieved an accuracy of 80%, and increased to 88% when self-reported pain scales were incorporated alongside imaging data in distinguishing patients with NSSI from HC. It highlighted amygdala-parahippocampus and inferior frontal gyrus (IFG)-insula connectivity as pivotal in NSSI-related pain processing. After incorporating imaging data of DC, the model's accuracy reached 74%, underscoring consistent neural connectivity patterns. The GAT model demonstrates high predictive accuracy for NSSI, enhanced by including self-reported pain scales.

Conclusions: Our proposed GAT model underscores the significance in the functional integration of limbic regions, paralimbic regions and IFG in NSSI pain processing. Our findings suggest altered pain processing as a key mechanism in NSSI, providing insights for potential neural modulation intervention strategies.

Keywords: Nonsuicidal self-injury; functional magnetic resonance imaging; graph attention networks; graph neural networks; interpretable machine learning; pain-processing neural circuits.

Figure 1.

An overview of our NSSI risk prediction pipeline with the pain-processing brain graph. (A) Rs-fMRI and DTI data are preprocessed and employed for the brain graph construction, utilizing the AAL 90 template and pain-related brain regions identified from a prior systematic review. (B) FC and FA matrices are extracted and utilized to construct pain-related brain graphs. (C) These graphs are then fed into our proposed interpretable GAT, which predicts (D) NSSI risks and generates the learned connectivity. Rs-fMRI, resting-state functional magnetic resonance imaging; DTI, diffusion tensor imaging; AAL, automated anatomical labeling; FC, functional connectivity; FA, fractional anisotropy; GAT, graph attention networks; IFGoperc, inferior frontal gyrus opercular part; IFGtriang, inferior frontal gyrus triangular part; ORBinf, inferior frontal gyrus orbital part; INS, insula; ACG, anterior cingulate gyrus; PHG, parahippocampal gyrus; AMYG, amygdala; THA, thalamus; STG, superior temporal gyrus; L, left; R, right.

Figure 2.

Interpretation of the learned results of the GAT model in the NSSI pain-processing network based on brain imaging data in distinguishing NSSI patients from healthy controls. (A) UMAP visualization of subject-level graph embedding for the NSSI pain-processing network. (B) Group-level node feature comparisons for each brain region in the NSSI pain-processing network, with no brain regions exhibiting significant group-level differences after FDR correction (p_FDR < 0.05). (C) Original FA connectivity and (D) learned connectivity of the NSSI pain-processing model displayed in a circular graph with a threshold > 0.05. Connectivity between brain regions in the NSSI pain-processing network is determined using attention weights derived from the learned GAT model. GAT, graph attention networks; NSSI, nonsuicidal self-injury; UMAP, uniform manifold approximation and projection; FA, fractional anisotropy; IFGoperc, inferior frontal gyrus opercular part; IFGtriang, inferior frontal gyrus triangular part; ORBinf, inferior frontal gyrus orbital part; INS, insula; ACG, anterior cingulate gyrus; PHG, parahippocampal gyrus; AMYG, amygdala; THA, thalamus; STG, superior temporal gyrus; L, left; R, right.

Figure 3.

Interpretation of the learned results of the GAT model in the NSSI pain-processing network based on both brain imaging data and self-reported pain scales in distinguishing NSSI patients from healthy controls. (A) UMAP visualization of subject-level graph embedding for the NSSI pain-processing network. (B) Group-level node feature comparisons for each brain region in the NSSI pain-processing network, with no brain regions exhibiting significant group-level differences after FDR correction (p_FDR < 0.05). (C) Original FA connectivity and (D) learned connectivity of the NSSI pain-processing model displayed in a circular graph with a threshold > 0.05. Connectivity between brain regions in the NSSI pain-processing network is calculated from the attention weights from the learned GAT model. GAT, graph attention networks; UMAP, uniform manifold approximation and projection; FA, fractional anisotropy; IFGoperc, inferior frontal gyrus opercular part; IFGtriang, inferior frontal gyrus triangular part; ORBinf, inferior frontal gyrus orbital part; INS, insula; ACG, anterior cingulate gyrus; PHG, parahippocampal gyrus; AMYG, amygdala; THA, thalamus; STG, superior temporal gyrus; L, left; R, right.