Using neurofeedback and machine learning to investigate the role of Beta oscillations in response inhibition

Abstract

Response inhibition is the ability to flexibly suppress unwanted or inappropriate actions, and it is a core component of healthy executive control. Impaired response inhibition is characteristic of a wide range of psychiatric diseases and of normal ageing. The underlying neural mechanisms of response inhibition have been researched intensively over the past two decades. It is hypothesised that the right inferior frontal cortex (rIFC) plays a key role by inhibiting the motor cortex via the basal ganglia and thalamus. It is thought that communication in this network is facilitated through Beta oscillations. While the vast majority of literature has investigated averaged Beta power, recent literature suggested that Beta oscillations are characterised by short-lasting, transient ?Beta bursts?, and only appears to change in sustained amplitude if it is averaged over many trials. However, the detailed functional roles of Beta oscillations and the involved anatomical structures in response inhibition are still unclear. Studies using bigger sample sizes as well as different experimental and analysis methods are necessary to identify generalisable findings which can meaningfully inform potential therapeutic interventions and impact their success. Thus, the overarching aim of this thesis was to advance the existing knowledge of the role of Beta oscillations in response inhibition by using methods designed to establish predictive and causal relationships. Empirical Chapter 1 aimed to advance recent findings about the spatial and temporal role of Beta bursts and averaged Beta power in human response inhibition by establishing Beta parameters that are predictive of response inhibition behaviour. As a first step, the Beta burst detection method was optimised using machine learning in order to maximise the sensitivity of further analyses. Next, a large human electroencephalography (EEG) Stop Signal Task (SST) dataset (n = 218) and machine learning were used to search a wide range of spatial and temporal features to (1) classify successful versus failed Stop trials and (2) to predict individual Stop Signal Reaction Times (SSRT). This was done using three features of Beta oscillations: Beta burst rate, Beta burst volume (duration x frequency span x amplitude) and averaged Beta power. The last and crucial step of this study was to validate the machine learning results on a large external human EEG SST dataset (n = 201) in order to establish true out-of-sample predictive Beta parameters of human response inhibition. Beta burst volume was significantly more predictive of successful and fast stopping than Beta burst rate and averaged Beta power. The classification model generalised to the external dataset, therefore establishing true out-of-sample prediction. In contrast, the internal validation results for predicting individual SSRT did not generalise to external data, therefore providing no evidence for prediction. In conclusion, Chapter 1 supported the emerging view that transient Beta bursts are a more accurate representation of oscillatory Beta activity in the brain and suggest a predictive role for Beta burst volume in human response inhibition. Empirical Chapter 2 aimed to further elucidate the role of Beta oscillations in response inhibition behaviour and whether these neural oscillations are causal or epiphenomenal to behaviour. While Chapter 1 was able to establish a predictive relationship between Beta oscillations and response inhibition behaviour by recording Beta oscillations from the scalp while participants were performing the SST, experimental manipulation of brain oscillations is necessary in order to specify a causal role for neural oscillations in behaviour. In Chapter 2, right frontal averaged Beta power was experimentally manipulated by training participants (n = 44) to intrinsically modulate their right frontal averaged Beta power using EEG-based neurofeedback in a brain-computer interface over 6 days. Response inhibition behaviour was measured prior and subsequent to neurofeedback training using the SST to assess whether right frontal Beta modulation impacted response inhibition behaviour. Right frontal Beta modulation was evident only during neurofeedback task performance but did not lead to offline alteration of Beta oscillations characteristics (including Beta bursts) at rest, nor to changes in subsequent response inhibition behaviour. Likewise, a control group (n = 38) who underwent training of right frontal averaged Alpha power did not exhibit behavioural changes. Therefore, the findings in Chapter 2 did not suggest a causal relationship of right frontal averaged Beta power and response inhibition behaviour, supporting the findings in Chapter 1. Empirical Chapter 3 investigated the direct neural implications of increased right frontal Beta power without incorporating behavioural measures. Chapter 1 and 2 suggested that behavioural measures like SSRT might not be sensitive enough to reflect changes in behaviour. This leaves open the possibility that neuroimaging techniques like functional magnetic imaging (fMRI) might have adequate sensitivity to reveal neural changes following manipulation of Beta oscillations. In Chapter 3, a subset of neurofeedback-trained participants (n = 20) from Chapter 2 took part in an additional fMRI session. fMRI and a psychophysiological interaction (PPI) analysis were used to investigate changes in functional brain connectivity within the fronto-basal ganglia-thalamic-motor cortex pathway during intrinsic right frontal Beta upregulation (n = 12). This study also included a control group who were trained to upregulate their right frontal Alpha rhythm (n = 8). The results showed decreased functional connectivity between the rIFC and the right thalamus while participants employed the same mental strategy as that used when attempting to upregulate right frontal Beta. No functional connectivity changes were found for the control group. Therefore, the results of Chapter 3 suggest that, even though no changes in behaviour were found in Chapter 2 following right frontal Beta upregulation, it still activated the neural pathway involved in response inhibition by inhibiting one of the key players in stimulating motor drive. In conclusion, this thesis provided valuable evidence that averaged tonic Beta power and Beta bursts are distinct phenomena, and that Beta bursts are a more accurate representation of the neural mechanisms underlying response inhibition behaviour. The current findings also provided insight into designing and using BCI-based neurofeedback training as a potential therapeutic intervention for populations with an existing deficit in response inhibition.