Novel ML-Based Algorithm for Detecting Seizures from Single-Channel EEG

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

1) The bibliographic revision of the Introduction is very complete. The subject of biomedical instruments to monitor seizures using machine learning was extensively covered.

2) Lines 345 to 359 are non-sense. These lines were probably imported from a generic manuscript template.

3) The paper consists in an application of XGBoost technique to detect seizures in EEG time-series data, just one channel. Much of the work consists in data treatment and selection, but I had no time to examine properly the results.

 

Author Response

We appreciate your time reviewing the manuscript. Please find below our response to the comments:

1) Thank you, and yes the extensive background was required to justify the need for this work.

2) Noted. Deleted from the revised manuscript.

3) The results show high performance (89% sensitivity) in detecting seizures, and the use of a single channel is essential to enable utilizing this method in single-channel portable EEG machines.

Reviewer 2 Report

Comments and Suggestions for Authors

Overall, this is an interesting paper that will be well received by the
target field.

Specific Comments

1. Introduction- although well written, the introduction is very long
and should be edited / reduced- some of the sections included could
easily be removed without impacting the narrative.

2. Methods: Well written and good detail (no suggested changes)

3. Results: Data is clear, but the legends only provide basic
information (and should be expanded). Fig 2- should be relabelled to
make 1 and 0 more meaningful, and sensitivity, specificity, NPV and PPV
could be calculated / included (instead of simply presenting the basic
TP/FP and FN/TN values and percentages). The results text is also
relatively brief and could be expanded to improve the narrative and
analysis- for example, the text includes the statement "achieved a high
sensitivity to epileptic events"- this should be expanded to include
values for sensitivity, specificity, NPV and PPV. They also need to
perform a kappa analysis to show agreement with correct diagnoses.
Sensitivity is included in Table 2, but this is presented in the
discussion and not the results (formatting issue). Importantly, data
reviewed in the methods is also missing- especially the univariate
analysis / bivariate analysis

4. Discussion: well written, well referenced and well presented (no
suggestions)

Author Response

We appreciate your time and effort in reviewing this article 1) The introduction was revised as advised. The changes are tracked in the updated manuscript file. 2) Thank you 3) Figure 2 has been revised as advised. The additional parameters describing the classifier performance have been added to the text under the Results section. The changes are tracked in the updated manuscript file. Regarding data review, as it is a publicly available dataset with other researchers employing it in their studies, we opted not to analyze it further for variability, but we are willing to do it if we are advised to do so. 4) Thank you

Reviewer 3 Report

Comments and Suggestions for Authors

line 47-52 Please add information on the current available system with the types of sensors and input as well.

 

line 83 Please add the information about the simultaneous recordings. Is it ECoG, MEG or something else?

 

line 319  >The used features are Kurtosis, Energy, and Differential Entropy.

Please define these terminologies used in your analysis

 

line 329 >the kurtosis and energy data were normalized, while the differential entropy data was standardized

Please indicate the reason why the kurtosis and energy was normalized unlike the differential entropy standardized.

 

line357  please add any approval code or number here.

 

At the end of the Discussion or the conclusion,

you might add future work that advance this research field,

in addition to the caveats that you had in the current study.

 

Comments on the Quality of English Language

English is good.

Author Response

1) Added to the introduction starting at line 50 as advised. These changes are tracked in the modified manuscript. 2) This segment was removed from the introduction upon the request to revise the introduction to make it more concise. 3) We defined the features in the Methods section as advised, and added the formulae used in the Appendix. 4) the kurtosis and energy data were normalized without assuming a certain distribution of the data, while the differential entropy data was standardized as its distribution can be assumed to be Gaussian. We added this clarification to the text in the revised manuscript. 5) a statement and the database link have been added to the "Data Availability Statement" at the end of the manuscript 6) A new section has been added at the end of the conclusion section.

Reviewer 4 Report

Comments and Suggestions for Authors

It is comfortable for patients, with a comfortable position near the ears, allowing all mobility and flexibility. It is a good fit for wearable devices in long-term continuous monitoring, especially when detecting unexpected seizures in real-time. 

However, there is a need to evaluate whether rapid and direct assessment of overall brain activity is sufficient and can be sufficient in the case of epilepsy and whether it is a "winning option" over limitations. 

All at the cost of limited information on spatial aspects of brain function. Seizures can be associated with almost any brain area: temporal, frontal, occipital, and central. They also differ from person to person. Also, in an individual, the nature (type) of his seizures may vary from time to time. I wonder if single-electrode EEG can identify them well (and not confound with artifacts or lead to false positives). Of course, multiple channels allow more detailed and accurate brain activity mapping with less chance of error.

Single-electrode EEG has the potential for rapid and timely seizure response because it has high sensitivity, which is critical in epilepsy monitoring to minimize false negatives and ensure that actual seizures are detected, but again... The model's effectiveness may not be specific to certain types of seizures, and its generalization to a wide range of seizure patterns may be challenging and cannot be misclassified. Crises may manifest themselves in various ways, patterns may occur in ways that are impossible to detect or at least not accessible, and a more diverse dataset may be needed.

Author Response

Thank you for your effort in reviewing the article. We agree with your assessment on monitoring seizures using single-channel devices and the limitations that are imposed on such an approach that may impact their applicability for seizure monitoring, Several studies have investigated the inclusion criteria and the types/locations of seizures that can reliably be monitored with near the ear EEG single channel monitoring systems, and these studies have been ongoing since the introduction of near the ear EEG recording in the early 2000s. This study aimed to develop a simple yet reliable algorithm to automate seizure classification task, which if successful can then be implemented in existing portable recording devices. The results have indicated that the classification performance of the proposed algorithm is comparable with more advanced approaches while retaining simplicity and low computational costs. We think the chosen dataset was sufficient to reach this conclusion, especially since it has been used in previous studies of classification tasks, and allowed us to do a fair comparison of performance. A continuation of this work can focus on performing comparison studies between standard clinical EEG monitoring and a portable single-channel EEG monitoring device that implements the proposed algorithm.

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

NA