Temporal analysis of global cortical communication of cognitive tasks in coarse EEG information is still challenging due to the underlying complex neural mechanisms. This study proposes an attention-based time-series deep learning framework that processes fMRI functional connectivity optimized quasi-stable frequency microstates for classifying distinct temporal cortical communications of the cognitive task. Seventy volunteers were subjected to visual target detection tasks, and their electroencephalogram (EEG) and functional MRI (fMRI) were acquired simultaneously. At first, the acquired EEG information was preprocessed and bandpass to delta, theta, alpha, beta, and gamma bands and then subjected to quasi-stable frequency-microstate estimation. Subsequently, time-series elicitation of each frequency microstates is optimized with graph theory measures of simultaneously eliciting fMRI functional connectivity between frontal, parietal, and temporal cortices. The distinct neural mechanisms associated with each optimized frequency-microstate were analyzed using microstate-informed fMRI. Finally, these optimized, quasi-stable frequency microstates were employed to train and validate the attention-based Long Short-Term Memory (LSTM) time-series architecture for classifying distinct temporal cortical communications of the target from other cognitive tasks. The temporal, sliding input sampling windows were chosen between 180 to 750 ms/segment based on the stability of transition probabilities of the optimized microstates. The results revealed 12 distinct frequency microstates capable of deciphering target detections' temporal cortical communications from other task engagements. Particularly, fMRI functional connectivity measures of target engagement were observed significantly correlated with the right-diagonal delta (r = 0.31), anterior–posterior theta (r = 0.35), left–right theta (r = − 0.32), alpha (r = − 0.31) microstates. Further, neuro-vascular information of microstate-informed fMRI analysis revealed the association of delta/theta and alpha/beta microstates with cortical communications and local neural processing, respectively. The classification accuracies of the attention-based LSTM were higher than the traditional LSTM architectures, particularly the frameworks that sampled the EEG data with a temporal width of 300 ms/segment. In conclusion, the study demonstrates reliable temporal classifications of global cortical communication of distinct tasks using an attention-based LSTM utilizing fMRI functional connectivity optimized quasi-stable frequency microstates.

中文翻译：

---

频率脑电图微状态的血流动力学功能连接优化使注意力 LSTM 框架能够对不同认知任务的不同时间皮层通信进行分类

由于潜在的复杂神经机制，对粗脑电图信息中认知任务的全局皮层通信的时间分析仍然具有挑战性。这项研究提出了一种基于注意力的时间序列深度学习框架，该框架处理功能磁共振成像功能连接优化的准稳定频率微状态，用于对认知任务的不同时间皮层通信进行分类。70 名志愿者接受了视觉目标检测任务，并同时获取了他们的脑电图（EEG）和功能性磁共振成像（fMRI）。首先，对获取的脑电图信息进行预处理并带通至 delta、theta、alpha、beta 和 gamma 频段，然后进行准稳定频率微观状态估计。随后，通过同时引出额叶、顶叶和颞叶皮层之间的 fMRI 功能连接的图论测量来优化每个频率微状态的时间序列引出。使用微状态信息功能磁共振成像分析与每个优化的频率微状态相关的不同神经机制。最后，这些优化的、准稳定的频率微状态被用来训练和验证基于注意力的长短期记忆（LSTM）时间序列架构，用于将目标的不同时间皮层通信与其他认知任务进行分类。根据优化微状态转移概率的稳定性，选择 180 至 750 ms/段之间的时间滑动输入采样窗口。结果揭示了 12 种不同频率的微状态，能够破译目标检测与其他任务参与的时间皮层通信。特别是，观察到目标参与的功能磁共振成像功能连接测量与右对角线δ（r = 0.31）、前后θ（r = 0.35）、左右θ（r = − 0.32）、α（r = − 0.31) 微观状态。此外，微状态信息功能磁共振成像分析的神经血管信息揭示了 delta/theta 和 alpha/beta 微状态分别与皮质通讯和局部神经处理的关联。基于注意力的 LSTM 的分类精度高于传统的 LSTM 架构，特别是采样时间宽度为 300 ms/segment 的 EEG 数据的框架。总之，该研究证明了使用基于注意力的 LSTM 利用 fMRI 功能连接优化的准稳定频率微状态对不同任务的全局皮层通信进行可靠的时间分类。