Neural mass model-based study of frontal-temporal theta oscillations in human subjects during the performance of a cognitive control task

Abstract

Cognitive control, the ability to rapidly shift one's attention and behavioral strategy in response to environmental changes, is often compromised across psychiatric disorders. One of the well-validated behavioral paradigms for tapping into the cognitive control circuits is a cognitive interference task, where subjects must suppress a natural response to follow a less intuitive rule. Slower response times on these tasks indicate difficulty exerting control to overcome response conflict. Conflict evokes robust electrophysiological signatures, such as theta (4-8 Hz) oscillations in the prefrontal cortex (PFC). However, the underlying neural mechanisms of conflict-evoked theta oscillations in the PFC are not clear. The objective of this work is to use a neural mass model (NMM) to find feasible cortical networks generating theta oscillations during conflict processing in human subjects. We used intracranial EEG (iEEG) recorded from dorsolateral PFC (dlPFC) and lateral temporal lobe (LTL) of human subjects with intractable epilepsy undergoing invasive monitoring, while they performed a multi-source interference task (MSIT). We used a dynamic causal modeling (DCM) framework to simulate dlPFC-LTL theta using a Jansen-Rit NMM. We found significant evidence for an LTL input into the dlPFC during the initial 500 ms of conflict processing compared to a bidirectional connection between the dlPFC and LTL. We conclude that a neural mass modeling framework can be used to elucidate candidate mechanisms of neural oscillations underlying conflict resolution in human subjects.

I. Introduction

Neuropsychiatric disorders are the number one cause of disability and health-related economic burden in the United States[1] and are commonly resistant to established therapies. Current therapies do not target the cause of these disorders: dysfunctional communication in distributed brain circuits[2], [3]., hence there is a need for a shift in treatment paradigm from trying to treat ill-specified symptoms such as mood[4], [5] to targeting a more objective, rigorously measurable cognitive function. One particularly promising construct is cognitive control, that is often compromised across mental disorders, including depression, addiction, schizophrenia, and obsessive-compulsive disorder[6]. Cognitive control is required for the flexible allocation of mental resources and is essential in the service of goal-directed behavior[7]-[10]. One aspect of cognitive control is response inhibition or conflict resolution which can be efficiently studied using a cognitive interference task. Neurophysiological, neuroimaging and lesion studies indicate the involvement of the prefrontal cortex (PFC)[10] specifically the medial and dorsolateral PFC (mPFC, dlPFC) and the anterior cingulate cortex (ACC) while processing conflict. Control and conflict evoke robust electrophysiologic signatures, especially theta (4-8 Hz) oscillations in the mPFC[11]-[13] and low-frequency synchrony among frontal and subcortical structures[6], [14]-[16]. Prior studies on neural mechanisms underlying cognitive control have focused on non-invasive imaging techniques such as functional magnetic resonance imaging and scalp EEG. Recently stereotactic electroencephalography (sEEG) recordings, used for surgical evaluation of epilepsy patients, has emerged as a technique for studying brain networks, offering unparalleled access to deep and peripheral structures in humans[17], [18]. In addition to the historically implicated regions such as the dlPFC and the ACC, we also found a significant involvement of the lateral temporal lobe (LTL) in generating low frequency oscillations in the recorded sEEG while processing higher cognitive load[19]. Most epilepsy patients have electrodes implanted in the PFC and temporal cortex both of which are indicated in cognitive functions and hence may reveal previously unknown aspects of brain circuits involved in cognitive control.

Empirical time-frequency network connectivity analysis provides us information about involved regions and oscillations, however they do not provide directionality of inter-region and intra-region information flow. Determining causality using measures such as Granger-causality and other information flow metrics assumes statistical models without any biophysical basis of the signal source and hence are difficult to correlate to neural dynamics. In this work, we determine prefrontal-temporal cortical connectivity during conflict resolution by modeling recorded sEEG from the PFC and LTL using neural population models that incorporate some of the biophysics of neuronal activity. Specifically, we used a DCM framework for the model estimation. DCM is a generic Bayesian framework for inferring hidden neuronal states from measured brain activity[20]. It provides posterior estimates of neurobiologically interpretable quantities such as the effective strength of synaptic connections among neuronal populations and their context dependent modulation.

In the following sections, we describe the modeling framework and a DCM framework for parameter estimation. Then we present our results and conclude with discussion and future direction.

II. Methods

A. Subjects and experimental paradigm

The sEEG was recorded from five epilepsy patients with pharmaco-resistant complex partial seizures who were undergoing invasive monitoring for surgical evaluation at the Massachusetts General Hospital (MGH). All participants were voluntarily enrolled after fully informed consent approved by local institutional review board at MGH. The electrode implants were solely made on clinical grounds and not tailored for research purposes. Participants performed the Multi-Source Interference Task (MSIT) [21] with simultaneous recordings of behavior and sEEG from both cortical and subcortical brain structures. MSIT trials consisted of three numbers between 0-3, two of which had the same value (Fig. 1A). Participants had to identify, via button press, the identity of the number that was unique, not its position. Each trial contained one of two levels of cognitive interference/conflict. Low conflict trials had the unique number in a position corresponding to its keyboard button, and flanking stimuli were always ‘0’, which is never a valid response. High conflict trials had the position of the unique number different from the keyboard position, requiring execution of a non-intuitive visuo-motor mapping (Simon effect). On high conflict trials, the non-unique numbers were valid responses (flanker effect). Each participant performed 1-3 sessions of MSIT. Each session consisted of multiple blocks of 32 or 64 trials, with brief rest periods in between blocks. sEEG was recorded from a montage of 8-18 bilaterally implanted depth electrodes. Electrodes were localized by using a volumetric image coregistration procedure. Using Freesurfer scripts (http://surfer.mnr.mgh.harvard.edu), the preoperative T1-weighted MRI (showing the brain anatomy) was aligned with a postoperative CT (showing electrode locations). Electrode coordinates were manually determined from the CT[22]. The electrodes were then mapped to standard cortical parcels/regions using an automatic, probabilistic labeling algorithm[23]. Intracranial recordings were made using a recording system with a sampling rate of 2 kHz (Neural Signal Processor, Blackrock Microsystems, US). At the time of acquisition, depth recordings were referenced to an EEG electrode placed on skin (either cervical vertebra 2 or Cz).

Fig1:

A: Schematic of MSIT, B: Brain regions from where sEEG was recorded with example theta band sEEG in boxes. C: Schematic of the Janset Rit Neural mass model and governing equations.

B. Neural Data Preprocessing

The sEEG were analysed using custom analysis code in MATLAB (Mathworks). To reduce the influence of volume conduction[24], sEEG were bipolar re-referenced by subtracting those recorded at consecutive electrode contacts on the same electrode shank. All sEEG data were decimated to 1000 Hz and de-meaned relative to the entire recording. 60 Hz line noise and its harmonics up to 200 Hz were removed by estimating noise signals through narrow bandpass filtering, then subtracting those filtered signals from the original raw signal. For each participant, we chose a channel pair spanning the dlPF) and one spanning the LTL which showed maximum evoked response during MSIT. We then band pass filtered the chosen sEEG to retain spectral power in the theta (4-8 Hz) band and normalized it with respect to a 0.5 second baseline period preceding image onset. We then considered 3 overlapping 1 second time windows: 0-1, 0.5-1.5- and 1-2-seconds following image onset. The theta response in these three time intervals were used for modeling. Following initial data preprocessing, it was ensured that the time points of data acquisition were the same from trial to trial (the file will not be compatible with SPM12 otherwise). Then the function spm_eeg_load was used to import the data into a file compatible with the SPM12 software. The following preprocessing steps were then executed on the SPM12-compatible file: 1) trials designated as bad were excluded from analysis, 2) data was cropped such that only the channels of interest were isolated for further analysis, 3) data was sorted such that low-conflict trials were first and high-conflict trials were second, 4) data was averaged for low-conflict trials and high-conflict trials.

C. Jansen-Rit Neural mass model estimation

The pre-processed sEEG data from two channel pairs representing dlPFC and LTL were used for model estimation using a DCM framework. We used an established Jansen-Rit (J-R) neural mass model (NMM) of coupled cortical columns[25] as the generative model of sEEG in the dlPFC-LTL network[26]. Each region was modeled as a J-R NMM (Fig 1C) which is comprised of three neural subpopulations: an excitatory pyramidal neuronal population, an excitatory and an inhibitory inter-neuron population. Each neuronal subpopulation is modeled by two blocks. The first block is an alpha function block that convolves with the average input action potential density of a neural population to produce an average output postsynaptic membrane potential (PSP). It can either be excitatory (h_e(t) = Aate^−at, t > 0) or inhibitory (h_i(t) = Bbte^−bt, t > 0). The parameters A and B determine the maximum amplitude of the excitatory and inhibitory PSP respectively, while the parameters a and b represent the average reciprocal time constant of the passive membrane and other spatially distributed delays in the dendritic network. The convolution operation can be written as a double differential equation which can be re-written as 2 differential equations (Equation1). The second block is a sigmoid function block (S(v) = v_m/[1 + e^r(v0−v)]) that transforms the average membrane potential (v) of a neuronal population into the average action potential density fired by the neuronal population. In this model a pyramidal population receives feedback from the inter-neuronal populations via internal connections (C1-C4 in Fig 1C). An external input (U) is included to model external sensory stimulation. A typical J-R NMM without any external inputs is characterized by the following 6 state equations:

$${\stackrel{.}{x}}_0=x_3,{\stackrel{.}{x}}_3=AaS(x_1-x_2)-2a{x}_3-a^2{x}_0$$ (1a)

$${\stackrel{.}{x}}_1=x_4,{\stackrel{.}{x}}_4=Aa{C}_{2}S(C_1{x}_0)-2a{x}_4-a^2{x}_1$$ (1b)

$${\stackrel{.}{x}}_2=x_5,{\stackrel{.}{x}}_5=Bb{C}_{4}S(C_3{x}_0)-2b{x}_5-b^2{x}_2$$ (1c)

The extrinsic (forward and backward connects between dlPFC and LTL) and intrinsic (J-R NMM internal excitatory and inhibitory connections, C₁ − C₄) network connections were estimated using a DCM framework implemented in SPM12[27]. The DCM estimation procedure begins by specifying all possible connections (models) that are of interest to be estimated. The procedure works by considering every permutation of connections to identify the most likely model to explain the data. As feasible network models, we considered 4 variants differing in the extrinsic and intrinsic connections (Fig 2B): 1) Both forward and backward connections between dlPFC and LTL, 2) Forward and backward connections along with intrinsic connections, 3) Only forward connection from dlPFC – LTL, 4) Only intrinsic connections.

Fig2:

A: Theta power evoked during low and high conflict trials in the dlPFC and LTL. Bars represent log ratio of theta power compared to baseline with the median indicated by red bar. B: Example recorded and simulated normalized theta evoked response in the dlPFC and LTL during low and high conflict trials. C: Log evidence of 4 models using a multi-subject fixed effects BMS across 3 time intervals.

The most feasible network architecture for each time interval was selected based on a Bayesian model selection (BMS) procedure[28]. BMS is an established procedure in statistics that rests on approximating the model evidence, i.e., the probability of the recorded sEEG evoked response, given a model. This exploits a Bayesian cost function (called relative Free Energy), which quantifies model evidence (i.e., the Free Energy is a score of model fits). Then, the most likely model (among a set of candidate models) is the one with the largest Free Energy. This is quantified in terms of the Bayes Factor (BF). This is a Bayesian analogue of the usual odds ratio and quantifies the probability that one out of many models has generated the observed sEEG data. A fixed effects BMS was performed across five subjects which assumes that subjects’ data arose from the same underlying model.

III. Results

We first chose channel pairs that spanned the dlPFC and LTL as determined from the electrode localization procedure. We then chose one channel pair per brain region that exhibited significantly higher theta evoked during high conflict compared to low conflict trials (Fig 2A, p<0.05). We then fit a Jansen Rit model to each subject using DCM implemented in SPM 12 as described in section II. Across 5 subjects, 3 time intervals and 4 connectivity model architectures, 60 models were generated (example simulated theta using best model fit in Fig 2B). These models were then used in a multi-subject fixed effects BMS to evaluate the most feasible model across subjects for each time interval. We found that the winning network model depended on the time interval with respect to MSIT image onset. As shown in Fig 2C, in the 1 second period following image onset, the model that had the highest evidence from recorded dlPFC-LTL theta was Model 2 which included a bidirectional flow of conflict related theta as well as intrinsic network connections. In the period of 0.5-1.5 second after image onset which also overlaps with the time when most responses occur, we found that model 3 was the one with highest evidence. This indicates that during this period, conflict related theta modulates the network connection from dlPFC to LTL as well as intrinsic network connections. In the 3rd time period of 1-2 seconds after image onset, model 4 representing only intrinsic network modulation was found to be the winning model which indicates that during this period representing the post decision making phase the dlPFC and LTL are no longer communicating. This analysis reveals a dynamic network modulation caused by high conflict following MSIT image onset.

IV. Conclusion

We show that a DCM framework using neural mass models can be used to elucidate candidate mechanisms and network interactions producing observed neural oscillations underlying conflict resolution in human subjects. In this work, we sought to build on past DCM analysis [29], which showed that models that receive inputs at both the PFC and LTL were preferable to models with only one input. As expected, the PFC-LTL network model did depend upon the time interval being analyzed. We found that the network structure changed from a bidirectional to PFC→LTL and finally to uncoupled regions as we progressed from initial conflict processing to response into post response period. The intrinsic neural population connections within each modelled region were however consistent across all the 3-time intervals considered. This indicates that conflict related theta oscillations are generated both within and across the dlPFC-LTL network with a dynamic network structure across the time intervals considered. This analysis was however constrained to the use of a Jansen-Rit neural mass model and did not explore connectivity schemes that include lateral connections. Future work will expand on exploring additional connectivity options of different neural models (Canonical Microcircuit) as well as using additional subject datasets and brain structures such as the cingulate. In conclusion, we have established a modeling framework for inferring network connections involved in cognitive processing by using biophysical inspired neural mass models as the generators of hidden neural dynamics. This can potentially help inform the target and time period for modulating conflict processing via neuromodulation interventions such as electrical stimulation.

Clinical Relevance—

This can be used to find feasible target mechanisms for designing therapy in patients with compromised cognitive control. This framework can also be expanded to serve as an in-silico test bed for designing and testing neuromodulatory interventions such as electrical stimulation for improving cognitive control in mood/anxiety disorders.