Summary
- There were issues with data quality due to different montages and preprocessing steps increasing variability.
- The machine learning classification was very accurate 97% at discriminating pre-meditation and endmeditation, indicating a difference between EEG co-spectra for these conditions.
- There was a relationship between time in meditation and probability of end-meditation classification, with D2S2 faster at inducing end-meditation state than D3S2.
- There were differences in the EEG power bands, with each meditation technique inducing different patterns of changes in the power bands.
Preprocessing Issues
Raw EEG data was processed, with some major inconsistencies found with respect to prior preprocessing:
- There were 3 different types of filtering previously applied
- No filtering
- 0.1-60 Hz band-pass
- 0.5-80 Hz band-pass
- However, in some cases, filtering did not look like it had been applied to all the electrodes
- There were different EEG electrode layouts (montages) used with different reference electrodes
- Some had more electrodes, most likely EMG sensors
- Some excluded A1 while others had included it
- Session length was quite variable with range of approximately ~6 minutes to 90 minutes
- This variability was not evenly spread across conditions with the majority of day 2 session 1 and day 3 session 1 recordings having a duration of ~6 minutes.
- This impedes the analysis of dynamic changes throughout meditation for these sessions and may make publishing such analysis difficult as this variability is a potential confounding variable.
Given the variability in preprocessing and montages, I selected the largest group: 0.5-80 Hz band-pass and excluded the other data sets from analysis.
Preprocessing
EEG data were exported in EDF format and imported into MNE-Python (Version 17.1, Gramfort et al., 2013, 2014) for subsequent analysis. The PREP pipeline was used detect channels that were corrupted by noise (Bigdely-Shamlo, Mullen, Kothe, Su, & Robbins, 2015) with all bad electrodes interpolated via the Spherical splines (Perrin, Pernier, Bertrand, & Echallier, 1989, 1990). The data were then bandpass filtered to 1-50-Hz with FIR filter (Rabiner & Gold, 1975). Potential Eye blinks were detected using a moving median, with a median between 30 to 300 microvolts with a window of 15 samples (60-ms) labelled as a blink, as measured at Fp1 and Fp2 electrodes. The data were transformed by the surface Laplacian (via spherical interpolation) to provide a more robust reference-free signal (Kayser & Tenke, 2006a, 2006b). The data surrounding the eye blink events were segmented int -500 to 500 ms epochs. Independent component analysis was conducted using the Picard algorithm (Ablin, Cardoso, & Gramfort, 2018a, 2018b), to isolate and removed EOG artifacts present in the data by selecting the component with the largest absolute Pearson r correlation coefficient to the eye blink epochs via find_bad_eog function in MNE-Python. The last five minutes of pre-meditation and meditation recordings were used to compare the effects of the various meditation types on EEG spectra, and the EEG recorded during meditation was used to assess the neural dynamics of meditation. Ggplot2 and MNE-Python were used to create of the Figures (Wickham, 2009).
Figure 1: Classifier Accuracy by Meditation Technique
References
Ablin, P., Cardoso, J. F., & Gramfort, A. (2018a). Faster ICA under Orthogonal Constraint. In ICASSP, ieee international conference on acoustics, speech and signal processing - proceedings. http://doi.org/10.1109/ICASSP.2018.8461662
Ablin, P., Cardoso, J. F., & Gramfort, A. (2018b). Faster independent component analysis by preconditioning with hessian approximations. IEEE Transactions on Signal Processing. http://doi.org/10.1109/TSP.2018. 2844203
Bigdely-Shamlo, N., Mullen, T., Kothe, C., Su, K.-M., & Robbins, K. a. (2015). The PREP pipeline: standardized preprocessing for large-scale EEG analysis. Frontiers in Neuroinformatics, 9 (June), 1–20. http://doi.org/10.3389/fninf.2015.00016
Carpenter, B., Gelman, A., Hoffman, M., Lee, D., Goodrich, B., Betancourt, M., . . . Riddell, A. (2017). Stan: A Probabilistic Programming Language. Journal of Statistical Software, 76 (1), 1–32. http://doi.org/10. 18637/jss.v076.i01