We recently adapted a real-time fMRI neurofeedback approach developed for studying attention in the normal brain [17] to attempt to alter the neurobiology underlying the negative attention bias (Figure 1). In a pilot feasibility study, participants with elevated depression were trained to selectively attend to an emotionally neutral target category (for example, scenes) for a period of time while ignoring an emotionally salient distractor category (for example, sad faces). All experimental parameters were identical to those reported by deBettencourt, et al. [15], including scanner make and model and scanning and experimental protocols. Further, all procedures were approved by the Institutional Review Board at the University of Texas at Austin and participants provided written informed consent.

Each training session in this study involved a series of scanning runs in two phases: a classifier-training phase and a testing/feedback phase. During the training phase, fMRI data were collected from participants as they performed a task requiring selective attention to a continuous stream of composite images containing overlaid (neutral) face-and-scene stimuli. Participants alternated between attending to the face or scene while trying to detect rare lure images. These data were used to train a pattern classifier to decode neural activity that reflected attention to face vs. attention to scenes.

During the testing/feedback phase, fMRI data were collected and decoded in real time using the trained classifier. Participants were always instructed to attend to scenes, and sad faces were introduced as distractors. The output of the classifier provided evidence about whether participants were attending to the correct category (that is, scene), and this was translated (within 2 s) into feedback for the participant. Feedback took the form of altering the visual display to encourage correctly directed attention and discourage incorrectly directed attention. For example, while the participants were supposed to be attending to scenes, if the classifier indicated that sad faces were distracting them, the proportion of the scene stimulus in the composite image was reduced (for example, from 50% scene/face to 30% scene/70% face).

This feedback served to ‘externalize’ participants’ attentional state, making their distraction by the sad faces more tangible. This also made the task of attending to scenes more difficult, providing an error signal that distraction was undesirable. The logic was that participants could learn from this tangible feedback about good and bad attentional states and gain an ability to better monitor and control these states. The alternative approach of making the scenes more visible when distraction by the faces occurred might have helped participants in that moment to reorient to the scenes; however, this would potentially incentivize lapses. That is, to simplify the task demands in this regime, the best strategy would be to seek distraction rather than avoid it. Ultimately, the effectiveness of different feedback regimes awaits further empirical study, but the approach used here of making the task more difficult when attention lapsed has proven effective in controls [15] and in depressed individuals, as shown below.

We ran a pilot study to demonstrate that this elaborate fMRI procedure is feasible in patients with depression. Seven adults with elevated symptoms of depression (mean Beck Depression Inventory-II [BDI-II] = 25; 4 female; mean age = 24) completed three sessions of neurofeedback training across a 5-day period, in between two laboratory assessment sessions. We were able to execute this procedure successfully, confirming the feasibility of the approach. Furthermore, the results were consistent with the possibility that this might be a useful approach. Specifically, improvements in attention control with training predicted improvements in mood symptoms across a 4-week follow-up period (Figure 2, left). Moreover, resting-state fMRI connectivity between frontal and parietal nodes of a previously identified attention control network [6] showed increased connectivity from before to after training (Figure 2, right).

These results must be interpreted with caution, as a control group was not included. Any future clinical study adopting this approach will need such a group, to ensure that the results cannot be attributed simply to practice with the task or other incidental aspects of the training. One control used in the previous study upon which this task was based [17] involved providing participants with sham feedback that was derived from other participants’ feedback sessions - and thus out of sync with their actual attentional state and thus presumably less useful for training. Future empirical work should include an appropriate active control condition.