Non-invasive real-time access to spatial attention information from 3T fMRI BOLD signals

Abstract

Access to higher cognitive functions in real-time remains very challenging, because these functions are internally driven and their assessment is based onto indirect measures. In addition, recent finding show that these functions are highly dynamic. Previous studies using intra-cortical recordings in monkeys, succeed to access the (x,y) position of covert spatial attention, in real-time, using classification methods applied to monkey prefrontal multi-unit activity and local field potentials. In contrast, the direct access to attention with non-invasive methods is limited to predicting the attention localisation based on a quadrant classification. Here, we demonstrate the feasibility to track covert spatial attention localization using non-invasive fMRI BOLD signals, with an unprecedented spatial resolution. We further show that the errors produced by the decoder are not randomly distributed but concentrate on the locations neighbouring the cued location and that behavioral errors correlate with weaker decoding performance. Last, we also show that the voxels contributing to the decoder precisely match the visual retinotopic organization of the occipital cortex and that single trial access to attention is limited by the intrinsic dynamics of spatial attention. Taken together, these results open the way to the development of remediation and enhancement neurofeedback protocols targeting the attentional function.

Beyond the brain-computer interface: Decoding brain activity as a tool to understand neuronal mechanisms subtending cognition and behavior

“…The decoding output on the test trials makes it clear that this assumption is false not only on error trials (where one expects attention to be miss allocated) but also on correct trials, such that attention can be either close or far to the cued location, although on average, attention is positioned on the cued location ( Astrand et al, 2016 , 2020 ). These findings have been confirmed in humans using fMRI recordings ( Loriette et al, 2021 ). Indeed, when predicting the spatial orienting of attention from BOLD activity in the striate and extrastriate cortex, while maximum decoding accuracy is achieved for the spatial location that the subjects are requested to attend to, on a significant proportion of the trials, attention is actually localized around the instructed location, thus indicating that attention is not always anchored at the cued location.…”

“…When a label is not correctly classified, this is considered as a false negative for the targeted label, and as a false negative for the predicted label. For example, the study of confusion matrices when decoding covert attention using fMRI shows clear attentional biases toward the lower visual field or along the horizontal and vertical meridians such that decoding accuracy is up to 10% higher at these locations relative to other locations in the visual field, thus confirming attentional biased observed behaviorally ( Zenon et al, 2008 ; Zénon et al, 2009 ; Loriette et al, 2021 ). In another fMRI study, Kim et al (2019) studied how the brain visually encodes tactile intensities.…”

“…Numerous methods are available in order to perform feature selection, ranging from statistical test as ANOVA, to principal component analysis (PCA), mutual information maximization, searchlight and others ( Pedregosa et al, 2011 ; Abraham et al, 2014 ; Allefeld and Haynes, 2014 ; Cunningham and Yu, 2014 ; Padmanaban et al, 2018 ). For example, in fMRI, an ANOVA-based feature selection precisely reveals the cortical topography for covert visual attention to guide single trial fMRI-based spatial decoding of attention ( Loriette et al, 2021 ). Another type of feature selection, called searchlight, aims at searching for the most informative features and select them to train the decoder, by looking at how each feature separately contributes to improving the classification accuracy.…”

“…For instance, the lateralization of the locus of spatial attention has been decoded in humans using event related brain potentials (ERPs) extracted from scalp EEG recordings in humans in response to visual presentations ( Trachel et al, 2015 ; Thiery et al, 2016 ). In fMRI, covert visuospatial attention can be decoded with a high level of accuracy when four positions are encoded (80% of accuracy) ( Andersson et al, 2012 ) and 40% accuracy when eight positions are encoded ( Loriette et al, 2021 ). Machine learning has also been applied to train predictive speech models using fNIRS, obtaining a 75% of accuracy in classifying long speech segments from brain activity ( Liu and Ayaz, 2018 ).…”