A Spike in Entropy Precedes the Mismatch Negativity; Linking Entropy and Prediction Error

2469 

Abstract Submission 

Michael Angyus¹, Fernando Rosas², Pedro Mediano², Robin Carhart-Harris³

¹Imperial College London, Portland, OR, ²Imperial College London, London, London, ³Department of Neurology, Neuroscape, University of California, San Francisco, San Francisco, CA

Michael Angyus, MSc Neuroscience  
Imperial College London
Portland, OR

Fernando Rosas  
Imperial College London
London, London

Pedro Mediano  
Imperial College London
London, London

Robin Carhart-Harris  
Department of Neurology, Neuroscape, University of California, San Francisco
San Francisco, CA

Karl Friston's free energy principle and predictive coding frame the brain as a hierarchically structured prediction machine that is iteratively testing neural models to achieve the most energy efficient arrangement of neural activity for survival. One potential indicator of the inefficiency of neural activity is the information theoretical quantification of entropy. Entropy has become an umbrella term for several different approaches to analyzing neural activity, not all of which are as interchangeable as the terminology from statistical mechanics would suggest. Lempel-Ziv complexity (LZC) has been used to reliably distinguish between conscious states using brain imaging data from electroencephalography (EEG). At its information theoretical basis, LZC is an estimator of Shannon's entropy rate, which is a measure of predictability or surprise. Additionally, a corollary of the entropic brain hypothesis is that the unpredictability of neural activity is reflective of the brains inability to predict - and therefore suppress - bottom-up stimuli. This leads to the following conjecture: that an increase in 'richness' of conscious content is at least partly facilitated by a disruption in the prediction updating systems at the sensory level. While LZC has been used to examine global states of consciousness, it is still unknown whether increases in conscious activity of awake states are related to brain wide entropy increases of model prediction error.

This conjecture could be tested using the mismatch negativity (MMN) paradigm, which has been used to investigate model updating in the auditory system via an event related potential associated with a mismatch between expectation and stimuli. To our knowledge, research has not attempted to directly relate any measures of entropy to this neural adaptation process. A recently developed algorithm for estimating signal complexity via state space entropy rate (CSER) allows for entropy measurements at a high enough temporal resolution for the analysis of event related EEG data, enabling us to test the hypothesis that prediction error has an associated spike in complexity that is resolved by updating the predictive model.

Our analyses reveal a significant spike in entropy occurring prior to the onset of the MMN, providing a first direct link between computational prediction error and neural activity associated with prediction error. Mixed effects regression revealed that the magnitude of this entropy spike was associated with strength of the MMN (Adjusted R2 = 0.385, ß = 0.012, p < 0.001). Mixed effects regression was also used to see if baseline entropy (prior to a given trial) was related to MMN magnitude, finding a significant positive relationship (Adjusted R2 = 0.239, ß = 0.023, p < 0.001).

The association between entropy and prediction error suggests brain entropy as a potential biomarker of the inability to suppress incoming stimuli. This examination of a low-level sensory model updating process provides first of its kind evidence for the interpretation of global entropy increases as a summation of various changing models across the brain. These results suggest that entropy is increased during the model updating process, providing support for predictive coding and the free energy principle.

Psychiatric (eg. Depression, Anxiety, Schizophrenia)

Attention: Auditory/Tactile/Motor

Consciousness and Awareness ¹

Perception: Auditory/ Vestibular ²

Sleep and Wakefulness

Computational Neuroscience

Consciousness

Electroencephaolography (EEG)

Learning

Perception

Plasticity

Seretonin

Sleep

Statistical Methods