Computational Modeling of Oddball Sequence Processing Exposes Common and Differential Auditory Network Changes in First-Episode Schizophrenia-Spectrum Disorders and Schizophrenia

Abstract

Background and Hypothesis

Differences in sound relevance filtering in schizophrenia are proposed to represent a key index of biological changes in brain function in the illness. This study featured a computational modeling approach to test the hypothesis that processing differences might already be evident in first-episode, becoming more pronounced in the established illness.

Study Design

Auditory event-related potentials to a typical oddball sequence (rare pitch deviations amongst regular sounds) were recorded from 90 persons with schizophrenia-spectrum disorders (40 first-episode schizophrenia-spectrum, 50 established illness) and age-matched healthy controls. The data were analyzed using dynamic causal modeling to identify the changes in effective connectivity that best explained group differences.

Study Results

Group differences were linked to intrinsic (within brain region) connectivity changes. In activity-dependent measures these were restricted to the left auditory cortex in first-episode schizophrenia-spectrum but were more widespread in the established illness. Modeling suggested that both established illness and first-episode schizophrenia-spectrum groups expressed significantly lower inhibition of inhibitory interneuron activity and altered gain on superficial pyramidal cells with the data indicative of differences in both putative N-methyl-d-aspartate glutamate receptor activity-dependent plasticity and classic neuromodulation.

Conclusions

The study provides further support for the notion that examining the ability to alter responsiveness to structured sound sequences in schizophrenia and first-episode schizophrenia-spectrum could be informative to uncovering the nature and progression of changes in brain function during the illness. Furthermore, modeling suggested that limited differences present at first-episode schizophrenia-spectrum may become more expansive with illness progression.

Introduction

Rapidly assessing salient sound change becomes impaired proximal to the onset of schizophrenia,^1–4 perhaps even prior to psychosis onset,¹ and the impairment progresses over time.⁵ Change detection relies on N-methyl-d-aspartate glutamate receptor (NMDAr) function,^6,7 a plasticity-related receptor implicated in schizophrenia.^8,9 Neurodevelopmental and neurodegenerative hypotheses¹⁰ of schizophrenia attempt to explain the biological basis for diagnostically significant changes in behavior/experience. Using neural activity generated by the auditory change detection system, we determined how multiple brain areas interacted to identify where anomalous processes occurred and how they differed in first-episode and established schizophrenia illness in an attempt to relate the auditory change deficits to cellular mechanisms and to determine whether deficits were affected by illness duration.

Many aspects of sound are patterned, such as the rustling of leaves, birdsong, the tick of a clock, or the whoosh of traffic. Once a sound pattern is known and predictable the sound carries little new information.¹¹ Automatic processes in the brain filter out predictable sound sources, freeing up resources for information that is relevant to whatever activity we are engaged in. This filtering function is enabled by building internal models of the causes of sensory input,^12,13 which dampen the neural response to predictable sounds. Importantly, the brain remains sensitive to sudden change in a predictable environment which elicits a larger neural response. This sensitivity to change ensures that events that do not match expectations are evaluated to determine if the deviation is an anomaly, a benign indication that circumstances have altered and the predictive model needs updating, or a threat.

Auditory change detection can be studied using a sequence of sounds and measuring how electroencephalography (EEG)-derived event-related potentials (ERPs) change in response to repetition and change. In schizophrenia, the reduction in responsiveness to repetitions^14,15 and the sensitization of response to deviations are impaired (for reviews^16–20). An index of change detection is constructed by subtracting the averaged response to repetitions from that to deviations, revealing the mismatch negativity (MMN), around 100–250 ms from sound onset.^21–23 Reduced MMN is recognized as a key index of biological changes occurring proximal to illness onset and deteriorating over the illness.^1,5,17

Internal models are built and modified over time to represent what is predicted (ie, the most likely sound properties and timing) and the trustworthiness of predictions (precision).^13,24,25 Theoretically, the activation to a sound is compared to anticipated activation, and any mismatch at any given level is communicated forward as an error signal. Backward connections from the higher (more rostral) network adjusts predictions to minimize the error signal. Although some aspects of model adjustment begin in the sub-cortical auditory pathway, cortical processing shows the greatest sensitivity to deviance detection²⁶ and engages a distributed network that includes primary, secondary, and pre-frontal regions of the brain.^25,27 In schizophrenia, the processes of modeling regularity (adaptation) and responding to a deviation (novelty detection) are grossly intact; the most likely sound properties are represented in internal models as inferred from reduced responsiveness to common regularities within sequences and sensitivity being evident in the factors that modulate deviant response amplitudes (such as probability and physical distinctiveness).¹⁷ It is the weighting of the predictions (precision) that appears to be affected as inferred from a reduced differential response between common and rare events. The deficit may be less affected at first-episode than in established illness.^19,20 However, how deficits develops over the course of disease progression is largely unknown, as are the underlying circuit abnormalities.

The timeframe over which patterns and precision are extracted, and the level of abstraction of the predictive model increases from more caudal to rostral nodes. Hence, an efficient and effective auditory change detection system engages a broad cortical network.²⁸ Within each cortical areas, the precision associated with a mismatch error signal is proposed to be expressed in the output gain of superficial pyramidal cells.^13,25 Several inhibitory and excitatory types contribute to gain, but MMN is critically dependent on short-term plasticity through NMDAr.^6,29,30 Pharmacological agents that act on NMDAr produce reliable modulation of MMN amplitude.¹⁷ The causes of vulnerability in this system in schizophrenia are not understood, but in-vivo^5,31 and postmortem^9,32,33 studies show structural changes in brain regions that form part of the auditory network. Furthermore, progressive structural changes, particularly in the left auditory cortex, correlate with progressive decline in pitch-deviant MMN amplitude over illness course^5,31 and MMN amplitude at first-episode also correlates with inferior frontal gyrus volumes.³¹ At a microcircuit level, postmortem studies show schizophrenia-related pathology in layer-specific supragranular pyramidal cells³³ and in parvalbumin fast-acting inhibitory interneurons.^34,35 These macro- and micro-structural differences may be related,³² with cell-type specific dendritic deficits leading to gray matter volume reduction, and pathophysiology as reflected in MMN. Further, it may be possible to identify subtle circuit anomalies at first-episode that become more pervasive with the illness that may not be as evident in the scalp-recorded MMN that merges activity from all contributing sources.

Responses to sound repetition and change obtained using scalp-recorded EEG can be modeled to identify how underlying brain regions and cell types interact to produce the sound-related activity. Here, dynamic causal modeling (DCM) was applied to standard and deviant responses from persons with first-episode schizophrenia-spectrum and individuals with established schizophrenia illness to identify how auditory network interactions are impaired and whether illness stage plays a role in the degree of deficits. It was hypothesized that DCM would reveal significant differences in the interactions that support auditory predictions and MMN in established illness, and that some differences would be evident at first-episode schizophrenia-spectrum selectively in left auditory cortex. The canonical microcircuit model³⁶ was used to detect anomalies in particular cell populations (eg, the superficial pyramidal cells and inhibitory inter-neurons) and whether differences in first-episode schizophrenia-spectrum were a subset of those in established illness.

Methods and Materials

Analysis of anterior scalp EEG voltage on the data was published previously.³⁷ Here a more sophisticated neural mass circuit-model analysis of multiple distributed sources and effective connectivity between sources (nodes) was applied with DCM. Forty first-episode psychosis schizophrenia-spectrum participants were compared with 40 matched healthy controls, and 50 established illness participants were compared to a matched group of 50 healthy controls (see Supplementary Methods for extended description).

Sound sequences comprised a standard tone (1 kHz, 50 ms duration, 5 ms rise/fall, 80 dB), a pitch-deviant (1.2 kHz, 50 ms duration, 5 ms rise/fall, 80 dB), and a duration-deviant (1 kHz, 100 ms duration, 5 ms rise/fall, 80 dB), presented with a stimulus onset asynchrony of 330 ms. A total of 1600 tones were presented, including 1280 standards (80%), 160 pitch-deviants (10%) and 160 duration deviants (10%). Due to time constraints, a subset of 6 first-episode, 13 first-episode matched controls, 14 established illness, and 16 established illness matched control participants were tested on only 800 tones. A choice was made to model the pitch-deviant MMN based on three considerations: firstly, the methods result in a large data set and results would be more straightforward and interpretable from a single deviant; secondly, the pitch-deviant is simpler to model because a pitch change induces less change in the morphology of the ERP³⁸; thirdly, in the pursuit of uncovering illness-related progression, the pitch deviant is a stronger choice while duration deviants are generally associated with reduced MMN from early in the illness, the MMN associated with a pitch change is less reduced at first-episode psychosis compared to established schizophrenia illness, but quite reliably observed later in the illness.¹⁶ While duration-deviant MMN shows smaller effect sizes at first-episode schizophrenia-spectrum than in long-term schizophrenia illness (d = ~0.4),^19,20 pitch-duration MMN is very little affected at first-episode schizophrenia-spectrum,³⁹ but shows progressive changes with psychosis duration, implying it is a biomarker of disease-related cortical change.⁵ We expected that the advanced discrete source and cortical circuit analyses would reveal subtle deficits not observed in the sensor-level compound electrical envelope of all activity.

EEG was recorded using a custom 72 channel Active2 high impedance system (BioSemi) cap. Sites included 70 10–10 scalp sites including left and right mastoids, the nose tip, and an electrode below the right eye. The bandpass was DC to 104 Hz (24 dB/octave roll off) digitized at 512 Hz. EEG was referenced to a common mode sense site (near PO1), with an active “driven-right leg” electrode on the homologous right site.

Auditory stimuli were presented while EEG was recorded and participants watched a silent video. Sensor space analysis and results are available in Supplementary materials where clear significantly smaller deviant tone responses were observed across fronto-central cortical sensor sites in the established schizophrenia illness group, but only on the trailing slope of the deviant response in the first-episode psychosis group. These results are similar to the more conventional peak amplitude findings at Fz reported in the original study³⁷ where the deviant-minus-standard MMN amplitudes were smaller in the established illness group comparison only.

The canonical microcircuit (CMC) version of DCM³⁶ was used to model the cortical source connectivity giving rise to the observed ERPs over 0–280 ms for the full electrode montage. The CMC model specifies not only the effective connectivity changes between cortical nodes, but also the effective connectivity between different cell populations in layers within each node.³⁶ The boundary element model in Montreal Neurological Institute space was used as per the default SPM12 template. The modeled network assumed six cortical sources identified as integral to auditory inference⁴⁰ (figure 1A with MNI locations as follows: left A1 [−42, −22,7], right A1 [46, −14, 8], left STG [−61, −32, 8], right STG [59, −25, 8], left IFG [−46, 20, 8], right IFG [46, 20, 8]). The eight models in figure 1A are differentiated by the inter-nodal coupling changes that could vary between standards and deviants; namely condition-specific “extrinsic” connectivity (forward and backward connections between distinct cortical sources) as well as condition-specific intrinsic coupling (gain parameters, that is, connectivity of neural populations within sources). Each node consists of a set of neural populations³⁶ as depicted in figure 1B. Differential equations model the dynamics of postsynaptic voltage and current in each neural population. Connectivity between nodes is reflected in both baseline (ie, common the standards and deviants) extrinsic connectivity (A parameters) and condition-specific (ie, standard versus deviant) changes (B parameters). Intrinsic connectivity is modeled by two types of parameters: (1) gain modulation on superficial pyramidal cells (modeled as self-inhibition and dependent on activity in other nodes, akin to NMDA-dependent short-term plasticity⁴¹), reflected in both the baseline (M parameters) and condition-specific effects (N parameters,⁴¹) and (2) the coupling between neural populations within each node (G parameters). Four of the latter connections were allowed to vary to capture the net inhibition/excitation differences present (see figure 1B). Finally, T parameters capture the time constants for each of the cell populations. DCM based on CMC has been used in several other studies of mismatch responses⁴¹ and validated using invasive recordings in animal models⁴² and humans,⁴³ as well as non-invasive recordings in humans. DCM applied to auditory-evoked mismatch responses has also shown convergence between models applied to invasive and non-invasive measurements.⁴⁴ While systematic simulation-based studies of parameter recovery have been so far largely limited to DCM for fMRI^45,46 (but see⁴⁷), both extrinsic and intrinsic parameters of DCM based on canonical microcircuits applied to EEG data have been shown to be reproducible based on real data⁴⁸ (ie, at realistic noise levels) and identifiable based on simulations.⁴⁹

Fig. 1.

(A) The eight discrete six-node models depicted with possible coupling changes identified. Each model included bilateral primary auditory cortex (A1), bilateral secondary auditory cortex (STG) and bilateral inferior frontal gyrus (IFG) with specific coordinates noted in text. The models were differentiated by which coupling changes were allowed to vary with detailed descriptions of forward (A, B) and intrinsic (M, N) connections provided in text. Each node within the model space was represented by a canonical microcircuit neural mass model (B). Within nodes G parameters capture intrinsic coupling to estimate the connectivity between superficial pyramidal cells, spiny stellate inter-neurons, inhibitory inter-neurons, and deep pyramidal cells contributing to group-specific changes in extrinsic connectivity between sources. The connections in green are those that were permitted to vary in the model. T parameters capture the time constants for different neuronal populations.

For DCM, the individual participant averaged standard and deviant waveforms (ERPs) were used to identify coupling changes within the model space that best accounted for the condition-specific differences (standard versus deviant). Bayesian model reduction was used following an inversion of the “full” model (figure 1A FBi) with SPM default priors, which incorporated changes in all identified parameters, to estimate model evidence for a range of “reduced” models where some parameters are not permitted to vary (the 8 models represented in figure 1A formed the reduced model space⁵⁰). The winning model was that defined by the higher posterior probability calculated using free-energy approximation to log-model evidence. As most subjects favored the same (FBi) model (see results) we then proceeded to explore the specific connections within that model that distinguished patients from controls.

In this second phase an iterative hierarchical implementation of the empirical Bayesian inversion method was used (Parametric Empirical Bayes or PEB⁵¹) to assess which combination of parameter changes best explained the group differences between the schizophrenia individuals and controls. The 64 models in the PEB included 63 combinations of A, B, G, M, N, T parameter types, plus the null model. Group-level effects were inferred by (re-)fitting the same model to each participant’s data under group constraints, with the prior updated to reflect the group average on each iteration (four iterations total). This process assumes that model parameters are normally distributed in the participant sample and updates the posterior distribution of the individual parameters (implemented using the SPM 12 function spm_dcm_peb_fit.m). Parameters with 99.95% Bayesian confidence intervals falling either side of zero (ie, corresponding to P < .005) were selected as statistically significant. Code for DCM analysis will be made available upon reasonable request.

Results

Group Differences in Coupling Strength Changes

The Bayesian model reduction resulted in the full model (figure 1A, FBi) being required to explain the deviant versus standard differences in 167 out of 180 participants (~93%), regardless of group membership (see figure 2C). This suggested the structure of the underlying DCM was unchanged between patients and controls and motivated an exploration of specific connection changes within the full model. The full model was then subjected to PEB analysis to determine which parameter changes were required to best model the group memberships with the first-episode schizophrenia-spectrum and established illness comparisons conducted separately (since, in SPM12 code for PEB, model search is implemented over two regressors—in our case, the constant/commonalities across groups and the difference between patients and controls).

Fig. 2.

A pictorial depiction of the results of the DCM analyses. The PEB analyses revealed that only intrinsic connections were required to model the patient versus control group differences for both the first-episode and established illness comparisons. (A) Depicts where M parameter coupling was significantly lower (blue arrows) or higher (red arrows) in the patient group relative to controls and the grey arrows represent where connections were modeled but did not vary between groups. (B) Depicts where G parameter coupling was significantly lower (blue) or higher (red) in the patient group relative to controls. The associated values for these comparisons are provided in Supplementary Table S1. (C) Details the winning model per participant 1–180 ordered first-episode control, established illness control, first-episode schizophrenia-spectrum, and established illness.

The patterns of significant difference in coupling strength changes between the patient and control groups are summarized in figure 2A for the nodal network and 2B for the CMC intrinsic connections. Recall that effects of standard/deviant tone probability in this analysis are modeled as condition-specific modulations of forward/backward extrinsic connections (B_BWD/B_FWD—B parameters) and intrinsic modulatory gain parameters (N parameters), while responses to tones generally were modeled as background modulations of forward/backward extrinsic connections (A_BWD/A_FWD—A parameters) and background intrinsic modulatory gain parameters (M parameters). The winning models for the first-episode schizophrenia-spectrum and established illness groups were very similar, with the former comparison requiring the M, G, and N parameters to account for group differences (posterior probability of 0.92), and the latter requiring only M and G parameters (posterior probability of 0.99).

Patient group differences in this analysis were therefore best modeled by intrinsic connectivity with change values provided in Supplementary Results Table S1, with first-episode psychosis patients showing an additional effect on mismatch-specific gain modulation. For the M and N parameters, the first-episode psychosis comparison revealed differences in the left A1 region only with higher precision overall (lower self-inhibition) in this area independent of tone probability and significantly lower deviant-specific precision (higher self-inhibition) in the same parameter in conditional values (ie, change in N parameters not shown in figure 2 but reported in Supplementary Results Table S1). The established illness comparison yielded the same baseline difference in self-coupling at left A1 but more widespread differences in the network with significantly lower precision-weighting on all tone responses at right STG and right IFG. In summary, the commonality in group differences from respective control groups was in the parameters reflecting activity-dependent gain modulation due to putative NMDA-mediated short-term plasticity scaled by neuronal inputs from other regions.

Both patient groups exhibited significant differences in coupling in G parameters relative to their comparison groups. The most consistent effects were evident in reduced self-modulation of the inhibitory inter-neurons and increased coupling between the inhibitory inter-neurons and the spiny stellate cells. The modulatory self-inhibition on inhibitory neurons was significantly lower in both patient groups relative to controls bilaterally in the IFG, and at left A1 in the first-episode schizophrenia-spectrum patients. The increase in inhibitory interneuron inhibition on spiny stellate cells (see figure 2B) was more pronounced in the patient groups relative to controls in several nodes. There were also widespread differences between patient and comparison groups in the modulatory self-inhibition on superficial pyramidal cells, but the patterns differed in the two groups. Increased modulation of this connection depicted by red arrows in figure 2B will result in increased self-inhibition consistent with a comparatively lower precision of prediction error responses to all tones. This is observed bilaterally in the IFG and right STG in first-episode psychosis patients, and in the right A1 in the established illness. In contrast, the same parameter is significantly lower in bilateral STG in the established schizophrenia group relative to the control group. The gain reflected in these self-connections is different than that reflected by the M and N parameters because it is activity-independent and akin to the influences of more classic neuromodulators⁴⁴ (eg, dopamine⁴⁵). Therefore, although this appears to be at odds with the results of the M parameters (where self-inhibition is higher overall in the established illness relative to controls at the right STG) this is not necessarily the case, since the two kinds of gain parameters can be putatively linked to different neuromodulatory systems.⁴¹

Discussion

It has long been proposed that smaller MMN responses to deviant sounds might provide useful information on some fundamental biological brain vulnerabilities in persons with schizophrenia.^52–56 Observations that MMN is dependent upon NMDAr function⁶ and that the measurable decline in MMN may appear proximal to the emergence of psychosis and diagnostically significant changes in function^2,57,58 provide further impetus to dig deeper into the underlying neural functions that differ in the schizophrenia illness. The present DCM results provide insight into which aspects of brain function involved in the generation of MMN might differ already at first-episode schizophrenia-spectrum and how this changes later in the established illness.

The group differences were accounted for by intrinsic (within area) effective connectivity differences rather than dysconnectivity between nodes in the network. The DCM analysis revealed that modeling the responses to standard and deviant tones and how they differ required effective connectivity changes throughout the full model space inclusive of forwards and backwards error passing between nodes. However, a much smaller set of parameter changes were required to differentiate between the patient and control groups implying that findings of smaller scalp-recorded MMN in traditional ERP studies are more likely to be due to differences in the weighting applied to error signaling rather than being linked to the error passing and updating through the network (A and B parameters). Both patient groups showed deficits in the activity-dependent (putatively NMDA-mediated) gain modulation at the left primary auditory cortex. Although MMN amplitude itself has been associated with structural brain changes in left Heschl’s and STG in schizophrenia,^5,31,59 the present findings model a consequent alteration in effective connectivity with the observation present in first-episode schizophrenia-spectrum and replicated in a separate group with established illness. This identifies left A1 as an early site of sound processing anomalies in the course of schizophrenia. In the present data a higher gain (lower self-inhibition) in response to all tones was present at left A1 and a comparatively lower gain for deviant over standard responses was present in the first-episode schizophrenia-spectrum individuals. Group differences were more widespread in the established illness and were evident in lower activity-dependent gain in right STG and right IFG to all sounds (ie, in upper levels of the network). Collectively these observations are consistent with an interpretation that disruption to typical precision-weighting on auditory responses is a key contributor to smaller MMN in schizophrenia with an early emergence of anomalous processing in left auditory cortex in first-episode psychosis. The lower activity-dependent gain on responses in right IFG and STG in the established illness may be influential in the more reliable and pronounced MMN reduction later in the illness.

Patient-related differences in G parameters were important to differentiating groups and were not more widespread in established illness. There were prevalent decreases in self-inhibition on inhibitory inter-neurons with a bilateral decrease at pre-frontal cortex common to first-episode schizophrenia-spectrum and established illness. Like differences in the left auditory cortex, the observation of the same significant parameter differences in the first-episode schizophrenia-spectrum and established illness individuals is essentially a within-study replication of a relevant coupling change anomaly and suggests an early emergence of this anomaly in disease. Self-inhibition of superficial pyramidal cells, an indicator of classic neuromodulation of gain, differed from controls within nodes at multiple levels. There were different patterns of gain difference where early illness was dominated by lower modulatory gain to all tones at IFG while established illness was associated with higher gain to all tones at STG.

Explanations for why some differences are expressed early versus later in the illness is speculative. The ability to produce large differential responses to deviant sound requires differential weighting or gain on output cells (the superficial pyramidal cells) of the responsive nodes. Different nodes within the network are theoretically linked to inferences over different timescales with sensory areas reflecting information gleaned over short timescales (100s of milliseconds) and higher levels reflecting information gleaned over many minutes or longer (eg, IFG).^28,60 The rostrocaudal gradient to timescales of pattern extraction could mean that the most efficacious modulations of responsiveness to sound will vary with sequence composition dependent on which patterns the system is attempting to predict.^60–63 The effect of errors detected over the short-term on the propensity to update an internal model of sound will depend on an estimate of how likely the sounds are to change estimated from patterning over a longer term.⁶⁴ Common anomalies in the left A1 in the patient groups could, for example, indicate differences in the tendency to adjust responsivity to sound properties over shorter timescales, with this moving to include differences in the ability to adjust responsiveness over longer timescales reflected in STG/IFG in the established illness. If this were the case, it is also possible that what is observed here as a down-weighting on responses at IFG in early illness in G parameters could reflect a shift to more local short timescale processing that later in the illness is better captured by higher weighting on response at STG.

There are relatively few DCM studies on MMN in schizophrenia and due to the differences in implementation it is not necessarily expected that results will converge on similar network anomalies. An early DCM study⁶⁵ indicated that bilateral A1 and right IFG were the sites required to explain group differences in data. Both persons with established schizophrenia and unaffected first-degree relatives differed in adjustments to excitability in superficial pyramidal cells in right IFG (left IFG not modeled) and like the present study, the differences were best modeled by intrinsic or within-area connections. More recently a version of CMC DCM was used to model altered auditory responses in schizophrenia and first-degree relatives to multiple sound stimulation paradigms.⁶⁶ For the oddball paradigm specifically, there are some similarities and differences with observations here. Firstly, in the present study most observed differences were condition (standard/deviant) independent which was not the case in this previous study where the schizophrenia group and relatives displayed a tendency for lower gain at IFG to deviants compared to standard. The results become more similar however, when controlling for medication dosage where the schizophrenia group showed higher superficial pyramidal self-inhibition in left and right IFG and reduced interneuron self-inhibition throughout. The former matches the observations in first-episode schizophrenia-spectrum individuals, and the latter is akin to both patient groups in the present study (figure 2B).

In summary, the present study provides further support for the notion that examining the ability to alter responsiveness to sound sequences in schizophrenia could be informative to uncovering the nature and progression of changes in brain function during the illness. The dense sensor array required to perform DCM analysis limits the ability to adopt the tool in routine assessments that typically opt for faster, reduced montage recordings. Nonetheless, the results suggest differences in the established illness can be modeled using DCM, with the results compatible with differences in both putative NMDAr activity-dependent plasticity and classic neuromodulation. Many of these differences are detectable in the first-episode schizophrenia-spectrum using DCM, despite minimal evidence of group differences in the scalp-recorded potential, and suggestive of common pathology that progresses or has progressive consequences over time. Relating these findings back to traditional MMN measures is somewhat complex as the average amplitude of MMN to a deviant tone is a singular index of the outcome of an underlying process. The present DCM analysis asks how averaged scalp-recorded potentials, and the differences between them, might be accounted for by the way that responsive brain regions (and cell types) affect change in responsiveness in other areas through incorporating what is known about biological underpinnings. As such the present results are based only on model fits using software designed to capture biologically plausible processes, but nonetheless these point to issues with computations that are reliant on observed regions and cell types implicated in schizophrenia.

The present results are based on ERPs collected from a traditional oddball sequence and among the implications is a likely role for altered inhibitory interneuron function in generating the smaller deviant responses that account for smaller MMN in schizophrenia. Different subtypes of interneuron are thought to convey different changes in responsivity with optogenetic suppression of parvalbumin fast-acting inhibitory interneurons leading to increased excitatory cortical responses (eg, like the findings here for M parameters at left A1).⁶⁷ Disruption to regular inhibitory interneuron activity in pre-frontal cortex in schizophrenia^68–71 is considered influential in working memory dysfunction⁷² and postmortem observations indicate pathology in layer-specific supragranular pyramidal cells³³ and in parvalbumin fast-acting inhibitory interneurons.^34,35 Indeed, the search for biomarkers of interneuron dysfunction has been posed as a priority for identifying those who might benefit from treatments with antioxidants.⁷³ These cell types feature prominently in parameters that captured the group differences in the present study and support continued exploration of how and why the auditory inference process underlying MMN generation is so sensitive to early signs of system vulnerability.

Funding

This work was supported by the National Health and Medical Research Council of Australia (APP2003933) and the US National Institute of Mental Health (R01 MH094328 and MH126951).

Disclosures

The authors have no conflicts of interest to declare.