Using real-time intracranial EEG and EEG-fMRI to investigate dynamic connectivity and epileptogenic activity in epilepsy disease

癫痫被越来越多地认为是一种异常的大尺度脑网络特性及其动力学导致癫痫事件的疾病[1,2]。此外，越来越多的计算和理论模型可以描述应该减少或增加癫痫脑活动的条件，这些模型可以映射到脑电图（EEG；在头皮或颅内获得）和功能性磁共振成像（fMRI） bbb的脑活动测量上。同时测量这两者在技术上具有挑战性，但它们一起提供高分辨率的空间和动态脑活动读数，并能够测量癫痫患者病理性脑动力学的周期，正如监督团队所证明的那样。我们有很多方法可以改变大脑网络活动的连通性和动态，包括经颅电刺激（TES）[4]，生物反馈[5]和认知任务[6]。然而，鉴于在非常大的参数空间背景下难以找到适当的个体特定参数，目前很少有工作解决优化这些方法的需要。自动神经科学家（AN）使用贝叶斯优化（作为由Leech主管开发的实时监督学习算法），其功能是通过大任务空间[7]进行闭环搜索。这种替代框架可以通过构建与实时分析相结合的神经适应范式来解决上述问题。该算法的特点是从样本空间中进行自动选择，并从中逐步学习，并可以利用其对空间的了解来优化后续选择。因此，在本研究中，将实时分析EEG-fMRI数据或eeg数据，以基于先前任务的实时结果迭代更新认知任务选择。这种方法比测试所有可能的任务要快，同时能够提供比随机测试更多的信息。它允许算法建立在其对功能组织的预先理解的基础上，通过迭代周期的测试和改进，在高维空间中产生一个健壮的模型，并优化任务建议，以获得最佳的大脑动力学。在此之前，这种方法已被用于最大化大脑中与认知过程相关区域的活动。在这里，目标（成本函数）将是癫痫活动的调节，例如在执行不同类型的认知任务时癫痫样放电的间歇率。

Epilepsy is increasingly understood as a disease where abnormal large-scale brain network properties and their dynamics are responsible for epileptic events[1,2]. In addition, there is increasing access to computational and theoretical models that can describe conditions that should reduce or increase epileptic brain activity, which can be mapped onto measures of brain activity from Electroencephalography (EEG; obtained on the scalp or intracranially) and functional Magnetic Resonance Imaging (fMRI) [3]. The measurement of both of these simultaneously is technically challenging, but together they provide a high resolutionspatial and dynamic readout of brain activity with the ability to measure periods of pathological brain dynamics in epilepsy as demonstrated by the supervisory team. e have a number of ways in which we can alter brain network activity in terms of connectivity and dynamics including transcranial electrical stimulation (TES) [4], biofeedback [5] and cognitive tasks [6]. However, there is currently very little work that addresses the need to optimise these approaches given the difficulty of finding appropriate parameters that are individual specific in the context of very large parameter spaces.The Automatic Neuroscientist (AN) uses Bayesian optimisation (as a real-time supervised learning algorithm developed by supervisor Leech) which functions on a closed loop search through a large task space [7]. This alternative framework might resolve the problem discussed above by constructing neuroadaptive paradigms combined with real-time analysis. The algorithm can be characterised by automatic selection from a sample space from which it progressively learns and can use its knowledge of the space to optimises the subsequent selection. In this study, the EEG-fMRI data or iEEG data would therefore be analysed in real time to iteratively update the cognitive task selection based upon the real time results from previous tasks. This approach is faster than testing all possible tasks while able to provide more information than simply testing at random. It allows the algorithm to build upon its pre-existing understanding of functional organisations by testing and refining in iterative cycles producing a robust model across a highly dimensional space and optimising task suggestion for optimal brain dynamics.Previously this approach has been used to maximise the activity in a brain region related to a cognitive process. Here, the target (cost function) would instead be a modulation of epileptic activity such as the rate of interictal epileptiform discharges as different types of cognitive tasks are performed.