Introduction

Real-time functional magnetic resonance (rt-fMRI) neurofeedback is a non-invasive therapeutic approach designed to provide feedback signal based on neural activity or connectivity to modulate brain dynamics and enhance associated cognitive processes ^1,2. Recently, Schenoist et al. ³ provided support for the feasibility of providing neurofeedback based on network strength of a large-scale distributed model of sustained attention. Extending the prior literature, we are interested in examining rt-fMRI neurofeedback to enhance attentional control in older adults. Aging is associated with significant cognitive decline especially on tasks of attentional control requiring focusing on particular sources of information while ignoring irrelevant information ^4,5. However, an important step in implementing rt-fMRI neurofeedback is to carefully select and optimize the target for providing neurofeedback, as modulation of target neural activity or connectivity will have important consequences for the cognitive domain of interest. As such, in this study, we examined differences in network strengths of two connectome-based models, derived in independent datasets, to predict sustained attention and mind-wandering, respectively.

Methods

Twenty-five older adults (aged 65-85 years) participated in a neuroimaging session at the Center for Cognitive and Behavioral Brain Imaging on a 3T Siemens MAGNETOM Prisma system using a 32-channel head coil. Participants were presented with the gradual continuous performance task (GradCPT ⁶) where they were shown gradually changing images of city and mountain scenes, and were asked to respond via a button press for city, but not mountain scenes. All participants completed two separate runs, each consisting of four 3-minute blocks interleaved with rest blocks.
To examine the connectome-based model that would be best suited for providing target signal during rt-fMRI neurofeedback, we applied two connectome-based network masks. The first mask – saCPM – was derived separately in a sample of 41 healthy older adults to predict d prime, or sensitivity, which takes both correct responses (i.e., hits) and errors of commission (i.e., false alarms) into account ⁷. Similarly, using the connectome-based predictive modeling approach, we also derived a mind-wandering connectome-based model to predict trial-to-trial fluctuation in reaction time during a task of sustained attention. Using data on 135 older adults from the Human Connectome in Aging (HCP-Aging ⁸) project, we derived a connectome-based model to predict reaction time variability (RT_CV= RT(SD)/RT(Mean)). As successful attentional control is likely a byproduct of the ability to select target representations, captured via d prime, and reduce interference associated with internal and external distractions, assessed in here via fluctuation in reaction time, we hypothesized that a mask derived from combining saCPM with mwCPM will likely serve as a potent target for neurofeedback. We thus applied the saCPM mask, including a high attention network and a low attention network, along with a mwCPM, including a high mind-wandering network and a low mind-wandering network, to the four blocks of the two gradCPT runs. For each of the four blocks in every run, we computed the network strength of the saCPM masks and the mwCPM masks and calculated the difference between the high and low networks. The difference score essentially represented a summary statistic indexing high connectivity in nodes associated positively with behavior and low connectivity in nodes associated negatively with behavior. For the combined mask, we combined high attention with low mind-wandering and low attention with high mind-wandering.

Results

Our saCPM combined model, indexing difference in network strengths between the high attention and low attention networks, showed an average difference of 0.19 (p<1e-30) between the network strengths of the two masks. This difference, as depicted in Figure 1, was seen in all blocks of runs 1 and 2, thus suggesting that the saCPM model differentiates between functional connections necessary for high attention from connections associated with poor performance. Interestingly, there was no block effect, such that across blocks there was no change in network strengths of the two maps. In contrast, the mwCPM, as depicted in Figure 2 had a smaller effect (~0.06); although there was a block effect such that the MW combined model showed a lower difference in Block 1 of runs 1 and 2 compared to the latter three blocks (p<0.022), suggesting that this model may be better suited to detect changes in mind-wandering over time. Our combined model, shown in Figure 3, showed both a block effect (P<0.04) as well as a statistically significant difference (average difference = 0.12; p < 1e-30) between the network strengths of functional connections representing high attentional control and low mind-wandering and low attentional control and high mind-wandering.

Discussion and Conclusions

Our results suggest that our two models derived in independent datasets are representative of high and low attentional states, thus making them appropriate targets for neurofeedback. The increase in network strength difference of the mwCPM model with time suggests that this model uniquely captures increasing mind-wandering in the gradCPT over time thus adding unique variance to the model. Our future directions include providing neurofeedback using the combined saCPM and mwCPM masks to increase the potential for neurofeedback training to enhance attentional control in older adults

Acknowledgements

No acknowledgement found.