Perception as Hierarchical Bayesian Inference - Toward non-invasive readouts of exteroceptive and interoceptive processing

Abstract

Psychiatry is stuck with symptom-based diagnostic schemes, which display poor predictive validity. Computational psychiatry aims at providing a more mechanistic description of mental disorders by using computational models to understand how (pathologies in) neural circuits produce (maladaptive) behavior. This thesis considers how hierarchical Bayesian models of perception can contribute to this endeavor by bridging the gap between computation and physiology in paradigms of exteroceptive and interoceptive processing. In the first two chapters, we examine the auditory mismatch negativity (MMN), an electrophysiological response to unexpected changes in the auditory domain, using electroencephalography (EEG), pharmacology, and predictions from a hierarchical Bayesian filtering model. We find that amplitudes of mismatch related EEG responses reflect precision-weighted prediction errors on two hierarchical levels of our model, pertaining to beliefs about statistical regularities, and their volatility, respectively. Our pharmacological results indicate that NMDA receptor function is crucial for volatility processing, and that mismatch responses in our paradigm are differentially sensitive to cholinergic (muscarinic) versus dopaminergic receptor status. These findings suggest that auditory mismatch paradigms might be capable of distinguishing among different disturbances in the neuromodulation of NMDA receptors, which have been suggested as pathophysiological pathways underlying symptoms of schizophrenia. Following this, we approach the mapping from computational quantities of our hierarchical Bayesian filtering model to (readouts of) neural activity more formally. Conceptualizing the model as a network of interconnected nodes, we derive predictions about the neural circuitry necessary to implement the message passing in this network, and about the time course of evoked responses that would ensue. In the last part, we acknowledge that mental and physical health are interdependent, emphasizing the need for formal accounts of body-brain interactions. We outline how the application of hierarchical Bayesian models to such interactions can provide a common taxonomy for psychiatry and psychosomatics. This taxonomy motivates the search for non-invasive readouts of the brain’s monitoring and regulation of bodily variables. As an example for such monitoring, we examine the heartbeat evoked potential (HEP), an EEG response to single heartbeats. We present evidence for increased HEP amplitudes during attentional focus to the cardiac domain, rendering the HEP a readout of the adaptive context-dependent up and down-regulation of interoceptive processing. Together, our results showcase how a hierarchical Bayesian approach to perception can guide the search for non-invasive readouts of interoceptive and exteroceptive processing. These readouts have the potential to stratify patients suffering from psychiatric and psychosomatic symptoms based on a mechanistic understanding of the underlying pathology that spans the computational and the physiological level.