Anterior insula coordinates hierarchical processing of tactile mismatch responses

Abstract

The body underlies our sense of self, emotion, and agency. Signals arising from the skin convey warmth, social touch, and the physical characteristics of external stimuli. Surprising or unexpected tactile sensations can herald events of motivational salience, including imminent threats (e.g., an insect bite) and hedonic rewards (e.g., a caressing touch). Awareness of such events is thought to depend upon the hierarchical integration of body-related mismatch responses by the anterior insula. To investigate this possibility, we measured brain activity using functional magnetic resonance imaging, while healthy participants performed a roving tactile oddball task. Mass-univariate analysis demonstrated robust activations in limbic, somatosensory, and prefrontal cortical areas previously implicated in tactile deviancy, body awareness, and cognitive control. Dynamic Causal Modelling revealed that unexpected stimuli increased the strength of forward connections along a caudal to rostral hierarchy-projecting from thalamic and somatosensory regions towards insula, cingulate and prefrontal cortices. Within this ascending flow of sensory information, the AIC was the only region to show increased backwards connectivity to the somatosensory cortex, augmenting a reciprocal exchange of neuronal signals. Further, participants who rated stimulus changes as easier to detect showed stronger modulation of descending PFC to AIC connections by deviance. These results suggest that the AIC coordinates hierarchical processing of tactile prediction error. They are interpreted in support of an embodied predictive coding model where AIC mediated body awareness is involved in anchoring a global neuronal workspace.

Fig. 1

Schematic depicting experimental setup and example stimulus train. Participants received mild somatosensory electrical stimulation (50 μs pulse) at twice sensory threshold on the median nerve of the left forearm. Subjective intensity was manipulated by switching between single pulse (bottom-row) and double pulse (top-row) trials. Double pulses were identical to single pulses, with a 100 ms interstimulus interval. Repetitions varied randomly from 3 to 7 standard stimuli before switching to the alternate stimulus type, with repetition counts randomly sampled from uniform multinomial distribution. The first stimulus of each train corresponded to a deviant (D), whereas the following repetitions were defined as standards (S1, S2, …, S6). For our fMRI analysis, we modelled the deviant trials and the third repetition as standard (see Methods for more details).

Fig. 2

Significant BOLD activations for the deviant > standard contrast. From left to right, images are centred on the peak voxel extracted for each region modelled in the DCM; dorso-posterior thalamus (panels A and B), anterior insula (C), middle cingulate (D), primary somatosensory cortex (E), and the middle frontal gyrus extending into DLPFC (F). Statistical parametric maps, family-wise error corrected for multiple comparisons P_FWE < 0.05, shown on average of 152 1 mm-resolution anatomical scans, normalized to MNI space. Corresponding in-plane MNI coordinate is shown below each image. Colorbar shows T-values at each voxel.

Fig. 3

Post-hoc Bayesian model selection (panels A and B), winning model (panel C), and mean coupling parameter plots (panel D). (A) Top left panel depicts the range of log-posterior probability among all models examined. The top middle panel (B) shows the posterior probability for all tested models. Model 255 had the highest probability of 0.79. Model 128 was the next most probable with a posterior probability of 0.06, resulting in a Bayes factor of 13.17 for the full versus reduced model, corresponding to strong evidence that the full model was the best explanation for the measured data within the tested model space. (C) Depiction of the winning full model (Model 255, far right peak in Fig. 3B), grey circles indicate modulation by the Deviant > Standard contrast. (D) Bar plot depicting mean posterior parameter estimates for all modulatory (DCM.Ep.B) parameters across participants, indicating the strength in Hertz with which each connection was modulated by deviant > standard stimuli. Error bars depict standard error. Modulations of inhibitory self-connections are shown at the right hand side of the graph.

Fig. 4

Full model and results of one-sample t-tests over estimated modulatory parameters. Red arrows depict results of one-sample t-tests over all 25 modulation parameters (inhibitory self-connections indicated by circular arrow around each region label). A general caudal to rostral flow increased effectivity connectivity in response to tactile deviants can be observed from thalamus (TH), and primary somatosensory cortex (S1), to anterior insula (AIC) and mid-cingulate (MCC), before reaching prefrontal cortex (middle frontal gyrus, MFG). In contrast to this feed-forward flow of modulatory influences, the AIC shows significant increases in both ‘forwards’ connections to cingulate and prefrontal cortex and ‘backwards’ connections with S1, indicative of error comparison. Interestingly, TH, AIC, and S1 self-connections are strongly disinhibited by tactile deviants. All p-values false discovery rate corrected for multiple comparisons, P_FDR < 0.05.

Fig. 5

Robust regression analysis with self-reported difficulty for detecting stimulus changes predicting the strength of deviance-driven modulation of effective connectivity. Left panel A, participants with enhanced modulation of the backwards MFG to AIC connection by surprising touch stimuli reported easier discrimination of stimulus changes. Data points depict individual participants; points shaded grey indicate those receiving down weights > 2 SD from the mean weighting (leverage points). These results suggest that top-down prefrontal to AIC connectivity underlies awareness of unexpected tactile changes. Right panel B depicts results of robust regressions (Tukey's biweight) over 20 extrinsic connection modulation parameters each predicting subjective detection difficulty, P_FDR threshold < 0.05.