Abstract
Neurofeedback requires a direct translation of neuronal brain activity to sensory input given to the user or subject. However, decoding certain states, e.g., mindfulness or wandering thoughts, from ongoing brain activity remains an unresolved problem.
In this study, we used magnetoencephalography (MEG) to acquire brain activity during mindfulness meditation and thought-inducing tasks mimicking wandering thoughts. We used a novel real-time feature extraction to decode the mindfulness, i.e., to discriminate it from the thought-inducing tasks. The key methodological novelty of our approach is usage of MEG power spectra and functional connectivity of independent components as features underlying mindfulness states. Performance was measured as the classification accuracy on a separate session but within the same subject.
We found that the spectral- and connectivity-based classification approaches allowed discriminating mindfulness and thought-inducing tasks with an accuracy around 60% compared to the 50% chance-level. Both classification approaches showed similar accuracy, although the connectivity approach slightly outperformed the spectral one in a few cases. Detailed analysis showed that the classification coefficients and the associated independent components were highly individual among subjects and a straightforward transfer of the coefficients over subjects provided near chance-level classification accuracy.
Thus, discriminating between mindfulness and wandering thoughts seems to be possible, although with limited accuracy, by machine learning, especially on the subject-level. Our hope is that the developed spectral- and connectivity-based decoding methods can be utilized in real-time neurofeedback to decode mindfulness states from ongoing neuronal activity, and hence, provide a basis for improved, individualized mindfulness training.
In this study, we used magnetoencephalography (MEG) to acquire brain activity during mindfulness meditation and thought-inducing tasks mimicking wandering thoughts. We used a novel real-time feature extraction to decode the mindfulness, i.e., to discriminate it from the thought-inducing tasks. The key methodological novelty of our approach is usage of MEG power spectra and functional connectivity of independent components as features underlying mindfulness states. Performance was measured as the classification accuracy on a separate session but within the same subject.
We found that the spectral- and connectivity-based classification approaches allowed discriminating mindfulness and thought-inducing tasks with an accuracy around 60% compared to the 50% chance-level. Both classification approaches showed similar accuracy, although the connectivity approach slightly outperformed the spectral one in a few cases. Detailed analysis showed that the classification coefficients and the associated independent components were highly individual among subjects and a straightforward transfer of the coefficients over subjects provided near chance-level classification accuracy.
Thus, discriminating between mindfulness and wandering thoughts seems to be possible, although with limited accuracy, by machine learning, especially on the subject-level. Our hope is that the developed spectral- and connectivity-based decoding methods can be utilized in real-time neurofeedback to decode mindfulness states from ongoing neuronal activity, and hence, provide a basis for improved, individualized mindfulness training.
Original language | English |
---|---|
Pages (from-to) | 565-574 |
Number of pages | 10 |
Journal | NeuroImage |
Volume | 185 |
Early online date | 11 Oct 2018 |
DOIs | |
Publication status | Published - 15 Jan 2019 |
Bibliographical note
© 2018 Elsevier Inc. All rights reserved.Keywords
- Neurofeedback
- Magnetoenecphalography
- Machine learning
- Mindfulness