Recurrent neural processing and somatosensory awareness

Abstract

The neural mechanisms of stimulus detection, despite extensive research, remain elusive. The recurrent processing hypothesis, a prominent theoretical account of perceptual awareness, states that, although stimuli might in principle evoke feedforward activity propagating through the visual cortex, stimuli that become consciously detected are further processed in feedforward-feedback loops established between cortical areas. To test this theory in the tactile modality, we applied dynamic causal modeling to electroencephalography (EEG) data acquired from humans in a somatosensory detection task. In the analysis of stimulation-induced event-related potentials (ERPs), we focused on model-based evidence for feedforward, feedback, and recurrent processing between primary and secondary somatosensory cortices. Bayesian model comparison revealed that, although early EEG components were well explained by both the feedforward and the recurrent models, the recurrent model outperformed the other models when later EEG segments were analyzed. Within the recurrent model, stimulus detection was characterized by a relatively early strength increase of the feedforward connection from primary to secondary somatosensory cortex (>80 ms). At longer latencies (>140 ms), also the feedback connection showed a detection-related strength increase. The modeling results on relative evidence between recurrent and feedforward model comparison support the hypothesis that the ERP responses from sensory areas arising after aware stimulus detection can be explained by increased recurrent processing within the somatosensory network in the later stages of stimulus processing.

Figure 1.

Localizer ERP and resulting dipoles. Somatosensory-evoked potential at electrode CP4 (grand-averaged localizer data) is shown in top right plot. Individual N20 component latencies were used to fit a single dipole in cSI [average MNI location (44, −21, 55), SD (5, 12, 8)]. Adding to this dipole, a symmetrical dipole pair in cSII/iSII [average MNI location (±50, −18, 21), SD (10, 7, 13)] was fitted to the 90–115 ms time window. Resulting dipoles are shown, plotted on the T1 single-subject template (MNI), for a subject with the smallest mean Euclidean distance of the three dipoles from average locations.

Figure 2.

A, N140 component distribution. EEG responses showed late divergence for detected and undetected stimuli peaking at 135 ms within a central contralateral cluster (second-level paired t tests, corrected for multiple comparisons using FDR at α_FDR = 0.05, q_FDR = 0.02 peak level). B, C, ERPs characterizing tactile detection. N140 amplitude was largest at electrode C6. A broadly distributed P300-like component peaked at electrode FCz. D, Topographical maps. Grand-averaged voltage distributions are shown on topographical maps for detected and missed targets at four latencies corresponding to typical somatosensory ERP components (N20, P60, N140, and P300, respectively).

Figure 3.

DCM. A, Structural DCMs. Individual dipole locations representing cSI and cSII/iSII were used to construct a structural model with three patterns of modulation by stimulus detection, representing feedforward, feedback, and recurrent processing between cSI and cSII. B, Model evidence. Models were compared using BMS for 11 data segments (plotted log-evidence difference against null model). C, Relative model evidence. Both recurrent and feedforward models explained data equally well for data segments <120 ms (log-evidence difference < 5). For longer data segments (>140 ms), the recurrent model was winning in all comparisons (very strong evidence > 5; dashed gray line). D, Connection strength across data segments. The feedforward connection (cSI–cSII, shown in blue) was stronger (100% posterior probability) after aware stimulus detection for data segments >80 ms (BMA, fixed effects). The feedback connection (cSII–cSI, shown in red) was stronger after detection for longer data epochs (>160 ms, 100% posterior probability), although the relative strength differences between conditions were less pronounced than for the feedforward connection. F, Forward; B, backward; det, detected; miss, missed. Asterisk marks significance.

Figure 4.

DCM output of the recurrent model. A, Source activity estimates. Summary of the mean estimated activity at dipole level in the winning (recurrent) model (n = 19). B, Comparison of data and DCM output. Grand-averaged ERPs (n = 19) at all channels for detected (left column) and missed (right column) stimuli, plotted as (1) scalp data (top row), (2) DCM-observed responses, projected from scalp data onto eight spatial modes accounting for the greatest amount of explained variance and backprojected into channel space (middle row), and (3) DCM-predicted responses of the recurrent model, backprojected into channel space (bottom row). L, Left; R, right; AF/F, anteriofrontal/frontal; FC/C, frontocentral/central; CP/P, centroparietal/parietal; PO/O, parietooccipital/occipital.