发掘精神分裂症大脑连接变异的集成聚类动态功能连接分析方法

A New Ensemble Clustering Dynamic Functional Connectivity Analysis Method for Exploring Brain Connectivity Variation in Schizophrenia

近年来,基于功能磁共振成像的动态功能连接在研究精神疾病方面表现出巨大的潜力。传统的基于聚类的动态功能连接分析方法(如K-means)受到类别个数、初始值和噪声的影响可能导致不可靠的功能连接状态(FCSs)。本研究提出了一种新的集成聚类动态功能连接分析方法。首先,利用多个不同的类别个数(k值)进行K-means产生多样性的簇;然后,基于Jaccard系数和随机游走挖掘不同簇之间的相似性以构造反应簇间关系的加权图;最终,对加权图进行社区检测获得可靠的元簇,并通过投票的方式将每个功能连接窗口分组到不同的元簇中,计算其质心作为功能连接状态。基于105名健康对照(HC)和70名精神分裂症(SZ)患者的fMRI数据,全方位比较所提出的方法和常用的基于K-means的方法进行动态功能网络分析的效果。相较于K-means方法,所提出方法在FCS 2上的平均类间相似性由83.2%降低至81.1%,在FCS 3上的平均类间相似性由76.8%降低至73.5%,聚类评估指标Davies Bouldin指数由6.74降低至6.44,Silhouette Coefficient指数由0.018提高至0.031。显示HC和SZ组的组间差异更集中于FCS 2,而K-means方法的组间差异分散在FCS 2、FCS 3和FCS 4。所提出方法可以自动获得FCSs数目,并具有更好的聚类质量和更可靠的功能连接状态,还发现了比K-means更有意义的组间差异,支持了该方法在探索精神疾病生物标志物方面的应用潜力。

In recent years,dynamic functional connectivity based on functional magnetic resonance imaging data has shown great potentials in studying psychiatric disorders.Traditional clustering-based dynamic functional connectivity analysis methods,such as K-means,are sensitive to the influence of class numbers,initial values,and noise,which may result in unreliable functional connectivity states(FCSs).This paper proposed a new dynamic functional connectivity analysis method based on ensemble clustering.First,K-means was performed with multiple different class numbers(k values)to generate diverse clusters.Then,the similarity between different clusters was mined based on Jaccard coefficient and random walk to construct a weighted graph reflecting the relationship between clusters.Finally,reliable meta-clusters were obtained by community detection on the weighted graph,and each functional connectivity window was grouped into different meta-clusters by voting,and its centroid was calculated as the functional connectivity states.Based on fMRI data from 105 healthy controls(HC)and 70 schizophrenia(SZ)patients,we comprehensively compared the performance of the proposed method with the commonly used K-means based method for dynamic functional network analysis.Compared with K-means method,the average inter class similarity of the proposed method decreased from 83.2% to 81.1% on FCS 2,and decreased from 76.8% to 73.5% on FCS 3.The Davies Bouldin index decreased from 6.74 to 6.44,and the Silhouette coefficacy index increased from 0.018 to 0.031.In terms of group differences,results obtained from our proposed method showed that the differences between the HC and SZ groups were mainly concentrated in FCS 2.In contrast,group differences resulted from K-means method were dispersed across FCS 2,FCS 3,and FCS 4,primarily due to subpar clustering metrics and high average inter-cluster similarity.In summary,the proposed method in this study was able to automatically determine the number of FCSs and exhibit better clustering quality and more reliable FCSs.Additionally,the method revealed more meaningful inter-group differences than k-means,indicating the potentials of exploring biomarkers for mental disorders.