Closed-loop training of attention with real-time brain imaging

Abstract

Lapses of attention can have negative consequences, including accidents and lost productivity. Here we used closed-loop neurofeedback to improve sustained attention abilities and reduce the frequency of lapses. During a sustained attention task, the focus of attention was monitored in real time with multivariate pattern analysis of whole-brain neuroimaging data. When indicators of an attentional lapse were detected in the brain, we gave human participants feedback by making the task more difficult. Behavioral performance improved after one training session, relative to control participants who received feedback from other participants' brains. This improvement was largest when feedback carried information from a frontoparietal attention network. A neural consequence of training was that the basal ganglia and ventral temporal cortex came to represent attentional states more distinctively. These findings suggest that attentional failures do not reflect an upper limit on cognitive potential and that attention can be trained with appropriate feedback about neural signals.

Figure 1. Real-time pipeline

(a) During feedback blocks, each brain volume (green) was acquired, preprocessed with masking, smoothing and z-scoring, and analyzed during the next volume with a multivariate classifier trained on volumes from recent stable blocks in which faces (blue) or scenes (pink) were attended. The result was averaged with the results for the two preceding volumes and used to update the stimulus shown to the participant on trials in the subsequent volume. (b) The classifier output indicated how attentive the participant was to the task-relevant versus task-irrelevant categories. This output was converted to a mixture proportion using a sigmoidal transfer function: less attention to the task-relevant category resulted in a decrease in the proportion of that category’s image in the composite stimulus on the next trial. These values were updated throughout the block as attention fluctuated over time.

Figure 2. Brain-behavior relationship

To verify that the classifier could provide useful feedback, we examined how predictive it was of behavior. (a) Across participants, average decoding accuracy from the stable blocks of the rtfMRI session (determined by offline MVPA with n-fold cross-validation) was highly correlated with behavioral performance in the pre-training session. (b) Within participants, there was greater classifier output for the task-relevant category than for the task-irrelevant category before correctly rejecting than before false alarming to a lure trial. Error bars represent ±1 s.e.m.

Figure 3. Change in behavior

Behavioral performance in the sustained attention task, as indexed by a non-parametric measure of sensitivity (A′), is plotted for the pre-training and post-training sessions. Participants who received accurate neurofeedback about their attentional state improved as a result of training, even though the feedback was no longer present in the post-training session. Control participants who received neurofeedback from other participants’ brains did not improve. A reliable group difference in improvement shows that accurate feedback boosted performance above and beyond practice effects and stimulus exposure. Error bars represent ±1 within-subject s.e.m.

Figure 4. Searchlight analyses

(a) Voxel-wise analyses were conducted to identify brain regions whose surrounding activity patterns for the two attentional states became more separable after neurofeedback training. We computed cross-validation accuracy for classifiers trained to decode face and scene attention from RT-residualized BOLD data using a sphere with a 1-voxel radius centered on each voxel. Increased separability was quantified as the difference in accuracy between the end (run n) and start (run 1) of the fMRI session. (b) A greater increase in classifier accuracy for the feedback group relative to the control group (P < 0.05, randomization test with threshold-free cluster correction; Montreal Neurological Institute (MNI) x, y, z coordinates in mm) was observed in left ventral temporal cortex (−34, −24, −25) and left basal ganglia (−18, −4, −5). Small clusters (not shown) were obtained in left lateral temporal cortex (−50, −45, −25; −51, −36, −20; −48, −42, −28) and left anterior temporal lobe (−26, 22, −32).

Figure 5. Potential sources of feedback

(a) Real-time whole-brain classifier output from the feedback blocks of a representative run for a single participant: evidence for each category (top) and evidence for the task-relevant minus task-irrelevant categories (bottom). (b) Offline classifier output for the same blocks from a perceptual network in occipitotemporal cortex (left) and an attentional network in frontoparietal cortex (right). The output from the whole-brain classifier was correlated with the outputs of the perceptual network classifier (r_wp) and attentional network classifier (r_wa) over time during the feedback blocks of each run. These correlations were averaged across runs within each participant to produce a measure of the extent to which the participant’s real-time feedback relied on information in each network. (c) This measure of reliance on each network was in turn correlated with the change in behavioral A′ from pre- to post-training to assess whether feedback from each network was useful for training.