Transfer Learning Algorithm of P300-EEG Signal Based on XDAWN Spatial Filter and Riemannian Geometry Classifier

Round 1

Reviewer 1 Report

Manuscript presents a form of dimensionality reduction followed by a classification algorithm for EEG signal. It is generally well structured, but some key points need to be addressed before manuscript can be considered for publication. 

Major comments:

The following comments about transfer learning are long, but likely easy to address. It seems manuscript would be improved with a formal definition of transfer learning as used in the described study. As of now, it is possible to get confused with transfer learning as used in the manuscript and test/validation data, cross validation, leave one out techniques commonly applied in other machine learning applications. The authors name "transfer learning" what seems to be more commonly referred to as cross validation scheme. Transfer learning is more commonly used for when a model is trained for a task A and repurposed to be used on a task B. Experiment 1: one-to-one transfer learning, data from one subject is left out and data from another subject is used for training. Experiment 2: all-to-one transfer learning, data from one subject is left out and the rest of the data is used for training. There seems to be no mixing between dataset I and II, which would more strongly characterize the methodology as transfer learning. As of now, Experiments 1 and 2 seem more like an ablation study regarding the amount of data needed - Experiment 2 outperforms Experiment 1 due to increased amount of data.

Now, there is the possibility that the authors are using a slighlty different naming convention, perhaps more common for the data/domain of the study. As the journal is multidisciplinary, if that is the case (convention to use "transfer learning" to simply imply data from one subject is used to classify data from a second subject from the same dataset - rather than completely different datasets) I strongly suggest references to the more common transfer learning applications (e.g., many applications of transfer learning for models trained on ImageNet to classify remotely sensed images, herbarya images, natural images, etc) and the more clear definition that, in this manuscript, transfer learning is used refering that model is trained on subject A and used to evaluate subject B, A and B coming from the same dataset. 

Authors infer that the "XDAWN spatial filter can effectively improve the quality of the evoked P300 components by considering the signal and noise simultaneously", however that affirmation seems not to be supported by information in the manuscript - there are no figures or any other data to support the claim. Moreover, there is no comparison of RGC with and without XDAWN or comparisons of improvement/deterioration of SVM and LDA when XDAWN data is used as input. Thus it is unclear what step, XDAWN or RGC, are the main responsible for classification performance. 

Minor comments:

Manuscript could be improved if read by a native English speaker. I highlighted some points in the minor comments below. Some of the citations are in the incorrect format. 

Line 18: please check if a space is missing between "interface" and "(BCI)"

Line 19: consider bein explicit on what "this" refers to in "to record this brain activity" to facilitate reading. 

Line 20: Imagining?

Line 23: check if a space is missing after comma: " signals,such"

Line 30: consider rephrasing to: the BCI system requires a long calibration
phase before each time it is used. 

Line  29 to 33: consider improving phrasing. What does "different subject" mean?

Line 37: Pieter-Jan et al. 

Line 37: proposed combining

Line 38: missing space "proportion(LLP)"

Line 38: Gayraud et al. 

Line 40: Lu et al. 

Line 43: Morioka et al. 

Line 45: One of the most potential method is the Riemannian geometry method

Line 47: SPD was not defined in the main text. 

Line 64 to 66: change "part will", "part is", to "part shows" or "part presents"

Line 70: "This" is used without a reference. Consider using "The dataset used in this study..."

Line 73: are "subject" and "patients" the same thing? If so, I suggest being consistent throughout the manuscript to facilitate reading. 

Lines 86-87: check misplaced comma. 

Lines 90-92: use active voice. 

Line 93: missing new lines. "XDAWN is a spatial filter. Its goal is to find" to "XDAWN is a spatial filter used to find". f is not defined. "d represents
the number of channels" of the ECG?

Line 100: "to directly mapping data" to "to directly map data" 

Line 110: "we mapping all" to "we map all"

Line 110: "and calculated" to "and calculate"

Line 112: Calculate should not be capitalized

Line 114:  sentence: "In [27],it used the reference matrix to do the affine transformation to solve this variability, but not consider the geometric structure of the covariance matrix" needs to be rephrased

Lines 117-125: Nice flowchart, it helps understand the process. Please review the description, it seems complete, but the main sentence is too long.

Line 127: "after flashing as features". What does that mean?

Line 144: LDA was not defined

Line 145: SVM was not defined

Table 5: dataset I?

Line 181: Needs capital letter. 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

The method presented here is still stable overall, but generally the increase in the amount of data has not brought a significant increase in performance.  There can likely be benefits made by the development to techniques that take advantage of feature space expansion.  For example testing the resulting covariance matrix with techniques like boosting where the geometry has been manipulated to strengthen discrimination power.

It is also suggested that the authors put some work into the visualization components.  Representation of the optimized covariance matrix computed from performance testing can be used as both training feature but also tools in the study of systematics.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

The authors did a good job reviewing the manuscript and answering all of the reviewers' questions and suggestions. I believe the manuscript is improved now and I do not have significant questions or suggestions for improvement at this moment. I recommend the manuscript for publication after some minor references and sentences corrections: 

References should not include authors first name. E.g., "Nathalie T. H. Gayraud et al. [14] complete a cross-session transfer..." should be ""Gayraud et al. [14] complete a cross-session transfer...". This is just one example, there are other occurrences of referencing with authors' first name. 

Line 110: "In [27], it tried to solve the cross-subject variability by doing the affine transformation of the covariance matrix, however, it do not consider the geometric structure of the covariance matrix. " sentence can be improved. What does 'it' refer to here? Could you use something like "In [27] Reuderink et al. tried to solve..., however their work did not consider..."?

Line 127: "We select the data from 0 - 0.5s after flashing as a sample". I now understand the meaning of the phrase based on authors comments. I suggest adding "In the experiment of the dataset I, rows and columns are flashed. In the experiment of the dataset II, characters are flashed. We select the data from 0 - 0.5s after flashing as a sample"

Author Response

Please see the attachment.

Author Response File: Author Response.pdf