A machine learning approach towards the differentiation between interoceptive and exteroceptive attention.

Interoception, the representation of the body's internal state, plays a central role in emotion, motivation and wellbeing. Interoceptive sensibility, the ability to engage in sustained interoceptive awareness, is particularly relevant for mental health but is exclusively measured via self-report, without methods for objective measurement. We used machine learning to classify interoceptive sensibility by contrasting using data from a randomized control trial of interoceptive training, with functional magnetic resonance imaging assessment before and after an 8-week intervention (N = 44 scans). The neuroimaging paradigm manipulated attention targets (breath vs. visual stimuli) and reporting demands (active reporting vs. passive monitoring). Machine learning achieved high accuracy in distinguishing between interoceptive and exteroceptive attention, both for within-session classification (~80% accuracy) and out-of-sample classification (~70% accuracy), revealing the reliability of the predictions. We then explored the classifier potential for 'reading out' mental states in a 3-min sustained interoceptive attention task. Participants were classified as actively engaged about half of the time, during which interoceptive training enhanced their ability to sustain interoceptive attention. These findings demonstrate that interoceptive and exteroceptive attention is distinguishable at the neural level; these classifiers may help to demarcate periods of interoceptive focus, with implications for developing an objective marker for interoceptive sensibility in mental health research.