Sandra Iglesias

Hierarchical Prediction Errors in Midbrain and Basal Forebrain during Sensory Learning

In Bayesian brain theories, hierarchically related prediction errors (PEs) play a central role for predicting sensory inputs and inferring their underlying causes, e.g., the probabilistic structure of the environment and its volatility. Notably, PEs at different hierarchical levels may be encoded by different neuromodulatory transmitters. Here, we tested this possibility in computational fMRI studies of audio-visual learning. Using a hierarchical Bayesian model, we found that low-level PEs about visual stimulus outcome were reflected by widespread activity in visual and supramodal areas but also in the midbrain. In contrast, high-level PEs about stimulus probabilities were encoded by the basal forebrain. These findings were replicated in two groups of healthy volunteers. While our fMRI measures do not reveal the exact neuron types activated in midbrain and basal forebrain, they suggest a dichotomy between neuromodulatory systems, linking dopamine to low-level PEs about stimulus outcome and acetylcholine to more abstract PEs about stimulus probabilities.

Translational Perspectives for Computational Neuroimaging.

Functional neuroimaging has made fundamental contributions to our understanding of brain function. It remains challenging, however, to translate these advances into diagnostic tools for psychiatry. Promising new avenues for translation are provided by computational modeling of neuroimaging data. This article reviews contemporary frameworks for computational neuroimaging, with a focus on forward models linking unobservable brain states to measurements. These approaches-biophysical network models, generative models, and model-based fMRI analyses of neuromodulation-strive to move beyond statistical characterizations and toward mechanistic explanations of neuroimaging data. Focusing on schizophrenia as a paradigmatic spectrum disease, we review applications of these models to psychiatric questions, identify methodological challenges, and highlight trends of convergence among computational neuroimaging approaches. We conclude by outlining a translational neuromodeling strategy, highlighting the importance of openly available datasets from prospective patient studies for evaluating the clinical utility of computational models.

Uncertainty in perception and the Hierarchical Gaussian Filter

In its full sense, perception rests on an agents model of how its sensory input comes about and the inferences it draws based on this model. These inferences are necessarily uncertain. Here, we illustrate how the Hierarchical Gaussian Filter (HGF) offers a principled and generic way to deal with the several forms that uncertainty in perception takes. The HGF is a recent derivation of one-step update equations from Bayesian principles that rests on a hierarchical generative model of the environment and its (in)stability. It is computationally highly efficient, allows for online estimates of hidden states, and has found numerous applications to experimental data from human subjects. In this paper, we generalize previous descriptions of the HGF and its account of perceptual uncertainty. First, we explicitly formulate the extension of the HGFs hierarchy to any number of levels; second, we discuss how various forms of uncertainty are accommodated by the minimization of variational free energy as encoded in the update equations; third, we combine the HGF with decision models and demonstrate the inversion of this combination; finally, we report a simulation study that compared four optimization methods for inverting the HGF/decision model combination at different noise levels. These four methods (Nelder–Mead simplex algorithm, Gaussian process-based global optimization, variational Bayes and Markov chain Monte Carlo sampling) all performed well even under considerable noise, with variational Bayes offering the best combination of efficiency and informativeness of inference. Our results demonstrate that the HGF provides a principled, flexible, and efficient—but at the same time intuitive—framework for the resolution of perceptual uncertainty in behaving agents.

The PhysIO toolbox for modeling physiological noise in fMRI data

BACKGROUND Physiological noise is one of the major confounds for fMRI. A common class of correction methods model noise from peripheral measures, such as ECGs or pneumatic belts. However, physiological noise correction has not emerged as a standard preprocessing step for fMRI data yet due to: (1) the varying data quality of physiological recordings, (2) non-standardized peripheral data formats and (3) the lack of full automatization of processing and modeling physiology, required for large-cohort studies. NEW METHODS We introduce the PhysIO Toolbox for preprocessing of physiological recordings and model-based noise correction. It implements a variety of noise models, such as RETROICOR, respiratory volume per time and heart rate variability responses (RVT/HRV). The toolbox covers all intermediate steps - from flexible read-in of data formats to GLM regressor/contrast creation - without any manual intervention. RESULTS We demonstrate the workflow of the toolbox and its functionality for datasets from different vendors, recording devices, field strengths and subject populations. Automatization of physiological noise correction and performance evaluation are reported in a group study (N=35). COMPARISON WITH EXISTING METHODS The PhysIO Toolbox reproduces physiological noise patterns and correction efficacy of previously implemented noise models. It increases modeling robustness by outperforming vendor-provided peak detection methods for physiological cycles. Finally, the toolbox offers an integrated framework with full automatization, including performance monitoring, and flexibility with respect to the input data. CONCLUSIONS Through its platform-independent Matlab implementation, open-source distribution, and modular structure, the PhysIO Toolbox renders physiological noise correction an accessible preprocessing step for fMRI data.

Multi-site reproducibility of prefrontal-hippocampal connectivity estimates by stochastic DCM.

This study examined the reproducibility of prefrontal-hippocampal connectivity estimates obtained by stochastic dynamic causal modeling (sDCM). 180 healthy subjects were measured by functional magnetic resonance imaging (fMRI) during a standard working memory N-Back task at three different sites (Mannheim, Bonn, Berlin; each with 60 participants). The reproducibility of regional activations in key regions for working memory (dorsolateral prefrontal cortex, DLPFC; hippocampal formation, HF) was evaluated using conjunction analyses across locations. These analyses showed consistent activation of right DLPFC and deactivation of left HF across all three different sites. The effective connectivity between DLPFC and HF was analyzed using a simple two-region sDCM. For each subject, we evaluated sixty-seven alternative sDCMs and compared their relative plausibility using Bayesian model selection (BMS). Across all locations, BMS consistently revealed the same winning model, with the 2-Back working memory condition as driving input to both DLPFC and HF and with a connection from DLPFC to HF. Statistical tests on the sDCM parameter estimates did not show any significant differences across the three sites. The consistency of both the BMS results and model parameter estimates indicates the reliability of sDCM in our paradigm. This provides a basis for future genetic and clinical studies using this approach.

Bayesian inference, dysconnectivity and neuromodulation in schizophrenia

This scientific commentary refers to ‘Estimating changing contexts in schizophrenia’, by Kaplan et al. (doi:10.1093/brain/aww095). The paper by Kaplan et al. in this issue of Brain addresses one of the most interesting questions in contemporary schizophrenia research: the role of uncertainty during perception (Kaplan et al. , 2016). Uncertainty enjoys much interest in schizophrenia research as it may provide a crucial link between core clinical symptoms of schizophrenia—aberrant perceptual inference (e.g. hallucinations) and abnormal beliefs (delusions)—and long-standing neurobiological findings that patients with schizophrenia display widespread alterations in structural and functional brain connectivity (dysconnectivity). These two cardinal features of schizophrenia have been integrated in disease theories, which have developed in three waves. A first influential proposal was that dysconnectivity in schizophrenia arises from abnormal regulation of NMDA receptor (NMDAR)-dependent transmission by neuromodulatory (dopaminergic and cholinergic) influences (Friston, 1998). Given the critical role of NMDARs for synaptic plasticity and myelination, this suggested that both neurodevelopmental aspects of schizophrenia ( cf . abnormal pruning of connections by altered experience-dependent plasticity) and structural dysconnectivity might arise from a primary disturbance of NMDAR-dependent plasticity due to aberrant neuromodulatory control. Second, these putatively abnormal NMDAR-neuromodulator interactions (NNI) were proposed to cause a central computational impairment in schizophrenia: abnormal hierarchical Bayesian inference in the cortex (Stephan et al. , 2006). This proposal was inspired by the notion that the brain constructs a hierarchical and probabilistic model of the world in order to infer the environmental causes of its sensory inputs (predictive coding), and by the increasingly discernible importance of NMDAR-neuromodulator interactions for implementing hierarchical Bayesian inference in the brain (Fig. 1). Under generic conditions, belief updates in Bayesian inference are driven by prediction errors (the difference between actual and predicted inputs) but, critically, weighted by how uncertain or precise both predictions and sensory inputs are. While …

S23. INTRODUCING COMPASS: COMPARING BRAIN ACTIVITY ACROSS PATIENTS WITH DIFFERENTIAL TREATMENT RESPONSE IN SCHIZOPHRENIA – AN OBSERVATIONAL STUDY

Abstract Background Present pharmacological treatment approaches in schizophrenia rest on “neuroleptic” drugs, all of which act as antagonists at dopamine D2/D3 receptors but additionally display major variability in their binding capacity to neurotransmitter receptors (Van Os & Kapur 2009). At present, the choice of any particular drug does not rest on any principled criteria: Once individual treatment has been started, therapeutic efficacy is monitored clinically, and a switch to a different drug is initiated when clear improvements remain absent after a few weeks. It is presently not possible to predict in advance which patients will respond well to a particular drug and who will experience little or no benefit (Case et al. 2011; Kapur et al. 2012). For instance, clozapine and olanzapine are often prescribed after other antipsychotics have shown to be ineffective in patients with schizophrenia or related disorders due to their pronounced side-effects. Both drugs, clozapine and olanzapine, share certain pharmacodynamic properties with comparatively low affinity towards dopamine D2-receptors, but very high affinity towards muscarinic receptors – a unique constellation that distinguishes them from other common antipsychotics. Importantly, previous studies have shown that a subgroup of schizophrenia patients might particularly benefit from these properties (Raedler et al. 2003, Scarr et al. 2009). Here, we present an ongoing observational study (COMPASS) which builds on these observations and addresses the question whether functional readouts of dopaminergic and muscarinic systems in individual patients could enable personalised treatment predictions. Guided by the dysconnection hypothesis of schizophrenia (Stephan et al., 2009), which postulates aberrant interactions between NMDA receptors and neuromodulators like dopamine/acetylcholine, the COMPASS study adopts a neuromodeling approach. The focus is on EEG/fMRI paradigms and computational models with empirically demonstrated sensitivity for altered function of NMDA, dopamine and muscarinic receptors, respectively. Methods To detect even small effect sizes, the study aims to recruit N=120 patients with schizophrenia who begin treatment with, switch to, or augment medication with olanzapine or clozapine. If possible, a replication sample (an additional N=120) will be recruited, too. Patients will be examined +/- 96h relative to treatment onset. Data acquisition encompasses the following measurements: Clinical interview, EEG (working memory, reward learning under volatility, auditory MMN under volatility, “resting”-state), MRI (optional; fMRI during auditory MMN under volatility, “resting”-state, and structural imaging), blood samples (genetic and biochemical analyses). After 2 and 8 weeks a clinical follow-up is conducted. Results The study is ongoing. Discussion The EEG/fMRI data will be analysed by computational models that infer functional states of glutamatergic, dopaminergic, and cholinergic systems (for review, Stephan et al. 2015). Model parameter estimates will serve as features in machine learning analyses of treatment prediction (Brodersen et al. 2014). If successful, this proof-of-concept study will lead to clinically useful tests for predicting the efficacy of clozapine/olanzapine prior to or during very early treatment. This could have a significant impact on clinical management as it would enable predicting, at an early stage, the therapeutic benefit for individual patients. Our neuromodeling approach to individual predictions may thus provide a principled basis for treatment decisions, help spare side-effects and enable informed switches in treatment strategy.

F157. HIERARCHICAL PREDICTION ERRORS DURING AUDITORY MISMATCH UNDER PHARMACOLOGICAL MANIPULATIONS: A COMPUTATIONAL SINGLE-TRIAL EEG ANALYSIS

Abstract Background A central theme of contemporary neuroscience is the notion that the brain embodies a generative model of its sensory inputs to infer on the underlying environmental causes, and that it uses hierarchical prediction errors (PEs) to continuously update this model. In two pharmacological EEG studies, we investigate trial-wise hierarchical PEs during the auditory mismatch negativity (MMN), an electrophysiological response to unexpected events, which depends on NMDA-receptor mediated plasticity and has repeatedly been shown to be reduced in schizophrenia. Methods Study1: Reanalysis of 64 channel EEG data from a previously published MMN study (Schmidt et al., 2012) using a placebo-controlled, within-subject design (N=19) to examine the effect of S-ketamine. Study2: 64 channel EEG data recorded during MMN (between subjects, double-blind, placebo-controlled design, N=73), to examine the effects of amisulpride and biperiden. Using the Hierarchical Gaussian Filter, a Bayesian learning model, we extracted trial-by-trial PE estimates on two hierarchical levels. These served as regressors in a GLM of trial-wise EEG signals at the sensor level. Results We find strong correlations of EEG with both PEs in both samples: lower-level PEs show effects early on (Study1: 133ms post-stimulus, Study2: 177ms), higher-level PEs later (Study1: 240ms, Study2: 450ms). The temporal order of these signatures thus mimics the hierarchical relationship of the PEs, as proposed by our computational model, where lower level beliefs need to be updated before learning can ensue on higher levels. Ketamine significantly reduced the representation of the higher-level PE in Study1. (Study2 has not been unblinded.) Discussion These studies present first evidence for hierarchical PEs during MMN and demonstrate that single-trial analyses guided by a computational model can distinguish different types (levels) of PEs, which are differentially linked to neuromodulators of demonstrated relevance for schizophrenia. Our analysis approach thus provides better mechanistic interpretability of pharmacological MMN studies, which will hopefully support the development of computational assays for diagnosis and treatment predictions in schizophrenia.

Inferring Effective Connectivity from fMRI Data

A formal understanding of processes that result from the interaction of multiple elements is hardly possible without mathematical models of system dynamics. This is important in neuroscience, particularly in neuroimaging, where inference on causal mechanisms in neural systems, for example, effective connectivity, requires a model-based approach. Here, we focus on a Bayesian framework for inferring effective connectivity from functional magnetic resonance imaging (fMRI) and dynamic causal modeling (DCM). DCM is a generative model of fMRI data which links hidden neural activity via a biophysical forward model to measured data. Bayesian inversion provides both the parameter distributions of the model parameters and (an approximation to) the model evidence; the latter provides a principled basis for model selection. Following a methodological discussion of DCM, we conclude with an outline of its potential use for clinical applications.