Dynamic sensitivity analysis: Defining personalised strategies to drive brain state transitions via whole brain modelling

Abstract

Traditionally, in neuroimaging, model-free analyses are used to find significant differences between brain states via signal detection theory. Depending on the a priori assumptions about the underlying data, different spatio-temporal features can be analysed. Alternatively, model-based techniques infer features from the data and compare significance from model parameters. However, to assess transitions from one brain state to another remains a challenge in current paradigms. Here, we introduce a "Dynamic Sensitivity Analysis" framework that quantifies transitions between brain states in terms of stimulation ability to rebalance spatio-temporal brain activity towards a target state such as healthy brain dynamics. In practice, it means building a whole-brain model fitted to the spatio-temporal description of brain dynamics, and applying systematic stimulations in-silico to assess the optimal strategy to drive brain dynamics towards a target state. Further, we show how Dynamic Sensitivity Analysis extends to various brain stimulation paradigms, ultimately contributing to improving the efficacy of personalised clinical interventions.

Keywords: Brain State; Brain stimulation; Deep Brain Stimulation, DBS; Magnetic Resonance Imaging, MRI; Non-Invasive Brain Stimulations, NIBS; Position Emission Tomography, PET; Probability Metastable Substates, PMS; Spatio-temporal dynamics; Transcranial Magnetic Stimulation, TMS; Transition Probability Matrix, TPM; Whole-brain models; diffusion Magnetic Resonance Imaging, dMRI; dynamic Functional Connectivity, dFC; functional Magnetic Resonance Imaging, fMRI; static Functional Connectivity, sFC; transcranial Alternating Current Stimulation, tACS; transcranial Direct Stimulation, tDCS; transcranial Electric Stimulation, tES; transcranial Random Noise Stimulation, tRNS.

Graphical abstract

Fig. 1

Conceptual Overview of Dynamic Sensitivity Analysis A) Descriptive Analysis. Traditional statistical analysis between empirical fMRI brain recordings of different study groups (example control and patient groups). Dynamics can be described by various features across space and time, for example Global Brain Connectivity, Functional Dynamics Spectrum or Probability Metastable Substates (PMS). B) Explanatory Analysis. Generative modelling approaches to describe the emergent dynamics of coupled dynamical units in the brain network. Network models can be adjusted to approximate spatio-temporal features of brain dynamics at the individual or group-level by tuning model parameters relating to various brain mechanisms, such as global/local coupling strength, gain modulation, stability of oscillations or excitatory/inhibitory balance. Relating to different brain mechanisms, these parameters can subsequently be statistically compared between conditions to obtain model-based features characteristic of each brain state. C) Predictive Analysis. The framework of Dynamic Sensitivity Analysis consists in the systematic investigation of the optimal strategy promoting a transition between distinct brain states. This framework provides a non-invasive means to evaluate the nonlinear response of distinct perturbation strategies aimed at promoting a transition from an aberrant brain state to an optimal and healthy brain state.

Fig. 2

Design Overview of Dynamic Sensitivity Analysis A) Experimental Analysis. fMRI signal is converted into a spatio-temporal description. Here, we focus on the Probability Metastable Substates (PMS) as a way to summarise brain dynamics across the spatial and temporal dimension. B) Model Fitting. Whole-brain models for optimal and aberrant dynamics are optimised to the PMS. C) Dynamic Sensitivity Analysis. An optimal transition to the target state is systematically explored by applying a perturbation protocol with varying parameters. D) Dynamic Sensitivity Analysis Evaluation. Varying perturbation sites, profiles, time durations and intensities are explored and evaluated for the optimal fit to the target description of PMS.

Fig. 3

Types of Perturbation Protocols. A) Perturbation Site: Stimulations to promote brain transitions can be applied at the global level to all regions of interest, at the mesoscopic level defined by functional systems or other maps defining heterogeneity in space such as neurotransmitter receptor maps, T1/T2 weighting or transcriptomics gradients or at the local level for individual brain regions. B) Perturbation Profile: Stimulations to promote brain transitions can reflect intrinsic (changes in local brain dynamics such as the bifurcation parameter in Hopf-model and the gain function in Wilson-Cowan model) and extrinsic effects (modelled via an additional term reflecting the stimulation). Stimulations can be noisy, oscillatory or constant. C) Perturbation Duration: Stimulations for the extrinsic perturbation profiles can reflect on-going, constant, pulse-based short-lived or periodic effects.

Fig. 4

Personalised in-silico brain model to predict therapeutic outcomes. The principles of Dynamic Sensitivity Analysis using whole-brain computational models open up for novel clinical intervention design. The schematic provides a pipeline for how to combine functional information about the brain spatio-temporal dynamics in a model to describe the aberrant brain state (leftmost). Once the model is optimised, exhaustive Dynamic Sensitivity Analysis can be performed to establish perturbation sites, profiles and/or durations for optimally rebalancing the aberrant dynamics towards a target brain state. Lastly, based on the optimal perturbation profile, clinical interventions can be suggested that are most compatible with the suggested in-silico intervention.