Threshold selection for ASR and how it can be used to denoise blink artifacts in Muse EEG
Published:
Threshold selection for ASR and how it can be used to denoise blink artifacts in Muse
Some thoughts on using ASR after bandpass filtering as a potential good pipeline for a low-density EEG system like Muse (4 ch). Developed for students to quick prototype with Muse for a class I am guest lecturing for. Unlike denser systems, we cannot do ICA on Low-density systems like muse. Rejecting eye blinks and infrequent bursts by visual inspection might be challenging for long recordings. An appropriate threshold can minimize these artifacts while preserving clean signals. I cannot over-emphasis this enough. Selecting the right threshold is critical here!!! The default parameter for ASR is very conservative and removes too much data. This is variable depending on your data distribution and the nature of the study. Always check the data and perform sanity checks and do not go by default parameters.
The figure demonstraets a section of EEG during an eyes closed and eyes open condition. Here, the blinks when present are removed by ASR whereas the section during eyes closed during which there are no blinks, the data is retained. It shows how a carefully selected threshold can remove blinks while not rejecting cleaner part (section without blinks are retained). 
Here it is very critical that we select the threshold well. For e.g. a conservative threshod of say 5 standard deviation distorts/ reconstructs even the cleaner part of the data 
A relatively high threshold on the other hand fails to remove the noisy component as well. Finding the right balance is key. Sometimes, there would not be enough components to reconstruct corrupt PC and ASR flattens the data. Make sure to remove these segments from later stage 
The default of value is never a good value. Make sure the value is ideal for your data, do empirical visual testing to select an ideal threshold for your data 
**An important point to keep in mind is that ASR is based on Principal component analysis. A concern is there is 4 ch only and so 4 PCs. The reconstruction will be lossy and not ideal. Thread is not intended to say the method is error-proof, but an option to consider & to emphasize not to go with default:)
