Machine learning-based characterization of dynamic brain functional network connectivity in patients with first-episode schizophrenia
10.3760/cma.j.cn113661-20241015-00330
- VernacularTitle:基于机器学习的首次发作精神分裂症患者的动态脑功能网络连接特征分析
- Author:
Pei LIU
1
;
Yangyang LIU
;
Ningning DING
;
Shuaiqi ZHANG
;
Zixuan LIU
;
Zhaoxi ZHONG
;
Yuchun LI
;
Haisan ZHANG
Author Information
1. 新乡医学院医学工程学院 河南省神经科学与控制工程技术研究中心，新乡 453003
- Publication Type:Journal Article
- Keywords:
Schizophrenia;
First-episode;
Dynamic functional network connectivity;
Independent component analysis;
Machine learning;
Resting-state functional magnetic
- From:
Chinese Journal of Psychiatry
2025;58(6):470-479
- CountryChina
- Language:Chinese
-
Abstract:
Objective:Using resting-state functional magnetic resonance imaging (rs-fMRI), we explored the changes in dynamic functional network connections (dFNC) in the brains of patients with first-episode schizophrenia (SZ) and evaluated the potential clinical value of dFNC changes in combination with a machine learning model.Methods:Clinical data of 50 patients with schizophrenia (schizophrenia group), 29 males and 21 females, aged 18-47 (28.3±7.2) years, who attended the psychiatric department of the Second Affiliated Hospital of Xinxiang Medical College from January 2022 to August 2023, were retrospectively included. In the same period, 50 healthy controls matched for age and education (healthy control group) were recruited, of which 24 were male and 26 were female, aged 18-48 (28.0±6.9) years. The rs-fMRI imaging data were acquired for each subject. The dFNC cluster analysis was performed based on independent component analysis, and the differences between groups with different state FNC matrices were statistically analyzed. The dataset samples were divided into a training set (35 SZ patients and 35 healthy controls) and a validation set (15 SZ patients and 15 healthy controls) in a 7∶3 ratio. A machine learning classification model was constructed based on the dFNC matri. The performance of the model for distinguishing between schizophrenia and healthy controls was assessed by five-fold cross-validation using accuracy (ACC), recall (REC), F1 score, and area under curve (AUC) metrics of the working characteristics of the subjects.Results:Five network functional connectivity states were obtained by dFNC cluster analysis. Patients with first SZ showed a wide range of high connectivity and low connectivity changes on the neural dynamic functional networks, as shown by increased dynamic connectivity within the visual network (VIS) in state 1 (weak connectivity); The dynamic connectivity between executive control network (ECN) and VIS, frontal parietal network (FPN) and VIS decreases at state 3 (strong connectivity); The dynamic connectivity between default mode network (DMN) and FPN, DMN and ventral attention network (VAN) decreases at state 4 (weak connectivity). The machine learning results show that the classification model constructed by the dFNC matrix combined with SVM in state 3 (strongly connected) in the validation set obtains the best classification results (ACC=0.938; REC=0.938; F1=0.937; AUC=0.984), and the overall average classification ACC of the five states reaches 0.751, and AUC reaches 0.784.Conclusion:Patients with first-episode SZ have some brain functional network connectivity abnormalities, and a machine learning model based on dFNC features has high classification performance in distinguishing first-episode SZ from HC.
