Machine learning analysis reveals aberrant dynamic changes in amplitude of low-frequency fluctuations among patients with retinal detachment

Yu Ji; Yuan-Yuan Wang; Qi Cheng; Wen-Wen Fu; Shui-Qin Huang; Pei-Pei Zhong; Xiao-Lin Chen; Ben-Liang Shu; Bin Wei; Qin-Yi Huang; Xiao-Rong Wu

doi:10.3389/fnins.2023.1227081

Machine learning analysis reveals aberrant dynamic changes in amplitude of low-frequency fluctuations among patients with retinal detachment

Front Neurosci. 2023 Jul 20:17:1227081. doi: 10.3389/fnins.2023.1227081. eCollection 2023.

Authors

Affiliations

¹ Department of Ophthalmology, The First Affiliated Hospital of Nanchang University, Nanchang, Jiangxi, China.
² Department of Radiology, The First Affiliated Hospital of Nanchang University, Nanchang, Jiangxi, China.

Abstract

Background: There is increasing evidence that patients with retinal detachment (RD) have aberrant brain activity. However, neuroimaging investigations remain focused on static changes in brain activity among RD patients. There is limited knowledge regarding the characteristics of dynamic brain activity in RD patients.

Aim: This study evaluated changes in dynamic brain activity among RD patients, using a dynamic amplitude of low-frequency fluctuation (dALFF), k-means clustering method and support vector machine (SVM) classification approach.

Methods: We investigated inter-group disparities of dALFF indices under three different time window sizes using resting-state functional magnetic resonance imaging (rs-fMRI) data from 23 RD patients and 24 demographically matched healthy controls (HCs). The k-means clustering method was performed to analyze specific dALFF states and related temporal properties. Additionally, we selected altered dALFF values under three distinct conditions as classification features for distinguishing RD patients from HCs using an SVM classifier.

Results: RD patients exhibited dynamic changes in local intrinsic indicators of brain activity. Compared with HCs, RD patients displayed increased dALFF in the bilateral middle frontal gyrus, left putamen (Putamen_L), left superior occipital gyrus (Occipital_Sup_L), left middle occipital gyrus (Occipital_Mid_L), right calcarine (Calcarine_R), right middle temporal gyrus (Temporal_Mid_R), and right inferior frontal gyrus (Frontal_Inf_Tri_R). Additionally, RD patients showed significantly decreased dALFF values in the right superior parietal gyrus (Parietal_Sup_R) and right paracentral lobule (Paracentral_Lobule_R) [two-tailed, voxel-level p < 0.05, Gaussian random field (GRF) correction, cluster-level p < 0.05]. For dALFF, we derived 3 or 4 states of ALFF that occurred repeatedly. There were differences in state distribution and state properties between RD and HC groups. The number of transitions between the dALFF states was higher in the RD group than in the HC group. Based on dALFF values in various brain regions, the overall accuracies of SVM classification were 97.87, 100, and 93.62% under three different time windows; area under the curve values were 0.99, 1.00, and 0.95, respectively. No correlation was found between hamilton anxiety (HAMA) scores and regional dALFF.

Conclusion: Our findings offer important insights concerning the neuropathology that underlies RD and provide robust evidence that dALFF, a local indicator of brain activity, may be useful for clinical diagnosis.

Keywords: brain region; dynamic amplitude of low-frequency fluctuation; k-means clustering method; resting-state functional magnetic resonance imaging; retinal detachment; sliding window; support vector machine.