Postersession 1
Poster #: 28
Topic: Computational models
Wednesday, Sep 9, 2015
1st floor

A computational single-trial analysis of MMN under ketamine

Lilian Aline Weber1, Andreea Oliviana Diaconescu1, Christoph Mathys2, André Schmidt3, Michael Kometer4, Franz Vollenweider4, & Klaas Enno Stephan5

1Institute for Biomedical Engineering, Translational Neuromodeling Unit, ETH & University of Zurich, Zurich, Switzerland
2Wellcome Trust Centre for Neuroimaging, University College London; Max Planck UCL Centre for Computational Psychiatry and Ageing Research, London; Translational Neuromodeling Unit, University of Zurich and ETH Zurich, London, United Kingdom
3University of Basel, Basel, Switzerland
4Neuropsychopharmacology and Brain Imaging, University Hospital of Psychiatry, University of Zurich, Switzerland
5Institute for Biomedical Engineering, Translational Neuromodeling Unit, ETH & University of Zurich; Wellcome Department of Neuroimaging, University College London, Switzerland

While the MMN is frequently interpreted as an index of surprise or prediction error (PE), the exact neuro-computational processes generating it are still subject to debate. Here, we present computational analyses of single-trial EEG data using data from a previous MMN study contrasting placebo against the NMDA-R antagonist S-ketamine. For an auditory roving paradigm, we propose that the observer infers on two (hidden) causes of the sensory input: a matrix of transition probabilities between tones of different frequencies, and the volatility of this matrix. Using a Bayesian model of perception and learning (Hierarchical Gaussian Filter), we understand learning as hierarchically-coupled predictions whose update dynamics are controlled by the next higher level. We hypothesized that update signals during learning manifest as trial-wise variations in EEG signal amplitude, where the MMN as a difference signal emerges from higher PEs on deviant trials compared to standard trials. In a multiple regression of the trialwise EEG signals of 14 participants, we find significant correlations of model estimates of PEs and surprise over fronto-temporal channels. Critically, the signalling of PEs is reduced under ketamine as compared to the placebo condition. These results support the notion that the deviant negativity component of the MMN encodes indices of PE and that neuronal mechanisms for encoding trial-wise PEs are sensitive to NMDA receptor antagonism. Our single-trial analysis allows for a far more fine-grained computational interpretation of the MMN than conventional ERP analyses, distinguishing between different types of PEs and surprise, and is easily applicable to other experimental designs.