基于交替隐马尔可夫模型的改进动态功能连接估计
Improved dynamic functional connectivity estimation with an alternating hidden Markov model.
作者信息
Long Zhiying, Liu Xuanping, Niu Yantong, Shang Huajie, Lu Hui, Zhang Junying, Yao Li
机构信息
School of Artificial Intelligence, Beijing Normal University, Beijing, 100875 China.
The State Key Laboratory of Cognitive Neuroscience and Learning & IDG/McGovern Institute for Brain Research, Beijing Normal University, Beijing, 100875 China.
出版信息
Cogn Neurodyn. 2023 Oct;17(5):1381-1398. doi: 10.1007/s11571-022-09874-3. Epub 2022 Nov 3.
UNLABELLED
Dynamic functional connectivity (DFC) analysis has been widely applied to functional magnetic resonance imaging (fMRI) data to reveal the time-varying functional interactions between brain regions. Although the sliding window (SW) method is popular for DFC analysis, the selection of window length is hard, and the temporal resolution is limited by the window length. The hidden Markov model (HMM) without the limitation of window length has been proven to be able to estimate time-varying brain states from fMRI data. However, HMM tends to be overfitted in DFC analysis of fMRI data because of the high spatial dimension and the limited sample size of fMRI data. In this study, we proposed an alternating HMM (aHMM) method that used the functional connectivity estimation of SW to initialize the covariance matrix of HMM and adopted an alternating HMM procedure to reduce the number of parameters during each optimization. The simulated and real fMRI resting data from the Human Connectome Projects showed that aHMM produced better robustness to noise, parameter number and sample size in DFC estimation than SW and HMM. For the real fMRI resting data of cerebral small vessel disease (CSVD), results of aHMM revealed that amnesia and mild cognitive impairment (aMCI) caused the CSVD with aMCI (CSVD-aMCI) group tended to spend more time on the brain state with overall weak connections and less time on the state with overall strong connections than the CSVD-controls. Moreover, CSVD-aMCI showed significantly lower connectivity amplitude and higher connectivity fluctuation than CSVD-control. In contrast, HMM did not detect intergroup differences of the connectivity amplitude and fluctuations and SW did not detect intergroup differences of connectivity fluctuations and fraction of time. The results further indicated that aHMM outperformed HMM and SW in detecting inter-group differences of temporal properties of DFC and connectivity fluctuations.
SUPPLEMENTARY INFORMATION
The online version contains supplementary material available at 10.1007/s11571-022-09874-3.
未标注
动态功能连接（DFC）分析已广泛应用于功能磁共振成像（fMRI）数据，以揭示脑区之间随时间变化的功能相互作用。虽然滑动窗口（SW）方法在DFC分析中很流行，但窗口长度的选择很困难，并且时间分辨率受窗口长度限制。无窗口长度限制的隐马尔可夫模型（HMM）已被证明能够从fMRI数据估计随时间变化的脑状态。然而，由于fMRI数据的高空间维度和有限样本量，HMM在fMRI数据的DFC分析中容易出现过拟合。在本研究中，我们提出了一种交替HMM（aHMM）方法，该方法使用SW的功能连接估计来初始化HMM的协方差矩阵，并采用交替HMM程序在每次优化期间减少参数数量。来自人类连接体项目的模拟和真实fMRI静息数据表明，在DFC估计中，aHMM比SW和HMM对噪声、参数数量和样本量具有更好的鲁棒性。对于脑小血管病（CSVD）的真实fMRI静息数据，aHMM结果显示，与CSVD对照组相比，失忆和轻度认知障碍（aMCI）导致的CSVD伴aMCI（CSVD-aMCI）组倾向于在脑状态中整体连接较弱的状态上花费更多时间，而在整体连接较强的状态上花费更少时间。此外，CSVD-aMCI的连接幅度显著低于CSVD对照组，连接波动高于CSVD对照组。相比之下，HMM未检测到连接幅度和波动的组间差异，SW未检测到连接波动和时间分数的组间差异。结果进一步表明，在检测DFC时间特性和连接波动的组间差异方面，aHMM优于HMM和SW。
补充信息
在线版本包含可在10.1007/s11571-022-09874-3获取的补充材料。
Zotero 插件