Machine learning analysis reveals abnormal functional network hubs in the primary angle-closure glaucoma patients

Abstract

Background: Primary angle-closure glaucoma (PACG) is a serious and irreversible blinding eye disease. Growing studies demonstrated that PACG patients were accompanied by vision and vision-related brain region changes. However, whether the whole-brain functional network hub changes occur in PACG patients remains unknown.

Purpose: The purpose of the study was to investigate the brain function network hub changes in PACG patients using the voxel-wise degree centrality (DC) method.

Materials and methods: Thirty-one PACG patients (21 male and 10 female) and 31 healthy controls (HCs) (21 male and 10 female) closely matched in age, sex, and education were enrolled in the study. The DC method was applied to investigate the brain function network hub changes in PACG patients. Moreover, the support vector machine (SVM) method was applied to distinguish PACG patients from HC patients.

Results: Compared with HC, PACG patients had significantly higher DC values in the right fusiform, left middle temporal gyrus, and left cerebelum_4_5. Meanwhile, PACG patients had significantly lower DC values in the right calcarine, right postcentral gyrus, left precuneus gyrus, and left postcentral gyrus. Furthermore, the SVM classification reaches a total accuracy of 72.58%, and the ROC curve of the SVM classifier has an AUC value of 0.85 (r = 0.25).

Conclusion: Our results showed that PACG patients showed widespread brain functional network hub dysfunction relative to the visual network, auditory network, default mode network, and cerebellum network, which might shed new light on the neural mechanism of optic atrophy in PACG patients.

Keywords: PACG; SVM; brain network; degree centrality; fMRI.

FIGURE 1

The spatial of the DC within the PACG and HC with different correlation thresholds (r = 0.15, 0.2, 0.25, 0.3). The group means DC maps of the PACG and HC [(A) (r = 0.15), (B) (r = 0.20), (C) (r = 0.25), and (D) (r = 0.30)]. PACG patients showed remarkably similar altered degree of centrality brain areas relative to healthy control in the different correlation thresholds (r = 0.15, 0.2, 0.25, 0.3) (FDR correction p < 0.001). DC, degree centrality; PACG, primary angle-closure glaucoma; HC, health control; L, left hemisphere; R, right hemisphere.

FIGURE 2

Comparison of the DC between the PACG and HC. (A) (r = 0.15), (B) (r = 0.20), (C) (r = 0.25), (D) (r = 0.30). Significant degree centrality differences were found between the two groups. The blue areas indicate lower degree of centrality values (two-tailed, voxel-level P < 0.01, GRF correction, cluster-level P < 0.05). GRF, Gaussian random field; L, left hemisphere; MOG, middle occipital gyrus; and R, right hemisphere.

FIGURE 3

Classification results using machine learning analysis based on DC values. Function values of two groups (class 1: PACG group; class 2: HC group); The ROC curve of the SVM classifier with an AUC value of 0.85 (r = 0.15). (A) Function values of two groups (class 1: PACG group; class 2: HC group); the ROC curve of the SVM classifier with an AUC value of 0.85 (r = 0.20). (B) Function values of two groups (class 1: PACG group; class 2: HC group); The ROC curve of the SVM classifier with an AUC value of 0.85 (r = 0.25). (C) Function values of two groups (class 1: PACG group; class 2: HC group); The ROC curve of the SVM classifier with an AUC value of 0.85 (r = 0.30) (D).