Simple Neurofeedback via Machine Learning: Challenges in real time multivariate assessment of meditation state

Summary

Meditation proficiency is hard to achieve without feedback because the mind easily wanders. EEG neurofeedback could help by providing real-time assessment. This work proposes a lightweight scheme using an autoencoder model trained on EEG features from long-term meditators. The model runs in real time on short data segments from a few channels, using reconstruction errors or latent variables as feedback parameters to measure meditation ability. However, results show that meditation states overlap substantially in multivariate EEG features and have prominent temporal dynamics, which simple one-class algorithms fail to capture. Multiple improvements to the autoencoder are described to address these issues and enable high-precision neurofeedback protocols.

Study at a glance

Abstract

Attaining proficiency in meditation is difficult, especially without feedback since the mind may be easily distracted with thoughts and only long term efforts see any impact. Self-regulation would be much more effective if provided real time assessment and this can be achieved through EEG neurofeedback. Therefore, this work proposes a scheme for assessing meditation-like state in real time from short EEG segments, using low computational settings. Signal processing techniques are used to extract features from long term meditation practitioners’ multichannel EEG data. An autoencoder model is then trained on these features such that the model can be run in real time. Its reconstruction errors or its latent variables are used to provide non typical feedback parameters which are used to establish an objective measure of meditation ability. Our approach is optimised to have lightweight architectures handling small blocks of data and can be conveniently used on low density EEG acquisition systems as it requires only a few channels. However, our experimental results suggest that the meditation state has substantial overlap even in terms of multivariate EEG features and show prominent temporal dynamics, both of which are not captured using simple one class algorithms. Being an extremely flexible one-class model, we have described multiple improvements to the proposed autoencoder model to address the above issues and develop simple yet high precision neurofeedback protocols.