The Turner Institute for Brain and Mental Health, School of Psychological Sciences, and Monash Biomedical Imaging, Monash University, Victoria 3168, Australia; School of Physics, University of Sydney, New South Wales, 2006 Australia.
School of Physics, University of Sydney, New South Wales, 2006 Australia.
Neuroimage. 2022 Aug 1;256:119051. doi: 10.1016/j.neuroimage.2022.119051. Epub 2022 Mar 8.
Large-scale dynamics of the brain are routinely modelled using systems of nonlinear dynamical equations that describe the evolution of population-level activity, with distinct neural populations often coupled according to an empirically measured structural connectivity matrix. This modelling approach has been used to generate insights into the neural underpinnings of spontaneous brain dynamics, as recorded with techniques such as resting state functional MRI (fMRI). In fMRI, researchers have many degrees of freedom in the way that they can process the data and recent evidence indicates that the choice of pre-processing steps can have a major effect on empirical estimates of functional connectivity. However, the potential influence of such variations on modelling results are seldom considered. Here we show, using three popular whole-brain dynamical models, that different choices during fMRI preprocessing can dramatically affect model fits and interpretations of findings. Critically, we show that the ability of these models to accurately capture patterns in fMRI dynamics is mostly driven by the degree to which they fit global signals rather than interesting sources of coordinated neural dynamics. We show that widespread deflections can arise from simple global synchronisation. We introduce a simple two-parameter model that captures these fluctuations and performs just as well as more complex, multi-parameter biophysical models. From our combined analyses of data and simulations, we describe benchmarks to evaluate model fit and validity. Although most models are not resilient to denoising, we show that relaxing the approximation of homogeneous neural populations by more explicitly modelling inter-regional effective connectivity can improve model accuracy at the expense of increased model complexity. Our results suggest that many complex biophysical models may be fitting relatively trivial properties of the data, and underscore a need for tighter integration between data quality assurance and model development.
大脑的大规模动力学通常使用描述群体水平活动演变的非线性动力系统方程来建模，不同的神经群体通常根据经验测量的结构连接矩阵进行耦合。这种建模方法已被用于深入了解自发脑动力学的神经基础，这些动力学是通过静息状态功能磁共振成像（fMRI）等技术记录的。在 fMRI 中，研究人员在处理数据的方式上有很多自由度，最近的证据表明，预处理步骤的选择可能对功能连接的经验估计产生重大影响。然而，这种变化对建模结果的潜在影响很少被考虑。在这里，我们使用三个流行的全脑动力学模型表明，fMRI 预处理过程中的不同选择会极大地影响模型拟合和对发现的解释。至关重要的是，我们表明，这些模型准确捕捉 fMRI 动力学模式的能力主要取决于它们拟合全局信号的程度，而不是协调神经动力学的有趣来源。我们表明，广泛的偏差可以由简单的全局同步产生。我们引入了一个简单的双参数模型，它可以捕获这些波动，并且与更复杂的多参数生物物理模型一样表现良好。通过对数据和模拟的综合分析，我们描述了评估模型拟合度和有效性的基准。尽管大多数模型对去噪不具有弹性，但我们表明，通过更明确地建模区域间有效连接来放松对同质神经群体的近似可以提高模型准确性，而代价是增加模型复杂性。我们的结果表明，许多复杂的生物物理模型可能正在拟合数据相对简单的性质，并强调需要在数据质量保证和模型开发之间进行更紧密的集成。
