Talk:Seizure detection

RECOMMENDATION:

The author is recommended to consider probabilistic approaches to early detection of Seizures. These techniques fuse the multivariate data using model of normality by modelling the underlying data density. A probabilistic threshold is obtained from the model of normality to detect the seizure events as deviations away from normality.

References [1] Stephen Roberts, Lionel Tarassenko, Probabilistic Resource Allocating Network for Novelty Detection, Neural Computation, March 1994, Vol 6, No 2, Pages 270-284. [2] P.E. McSharry, T. He, L.A. Smith and L. Tarassenko, "Linear and non-linear methods for automatic seizure detection in scalp electroencephalogram recordings", Journal of Medical and Biological Engineering and Computing, vol. 40, no. 4, pp. 447-461, July 2002.

Reviewer B:

This is a comprehensive article about the problems and pitfalls associated with an automated algorithmic detection of epileptic seizures in electrophysiological time series, written by one of the experts and pioneers in the field. The section on Challenges in Seizure Detection is particularly helpful.

I only have a few minor suggestions for improvement.

1st paragraph: A clear distinction between the fields of seizure detection and seizure prediction would be helpful.

Fig. 1 and 2: Axis labels area barely legible.

Fig. 2: Do all of the algorithmic detections correspond to actual seizures? If not, which of them are true positives and which are false positives?

Fig. 5: This figure should have double width so the scaling on the abscissa remains constant when the second half of the data is added.

Fig. 8: Same as Fig. 5.

The paragraph on SDA Performance Assessment is short and succinct, which is good given the overall length of the article. I am not sure, however, whether Figs. 10 and 11 can be easily understood without a lot more explanation. So I wonder if it might be more helpful to just omit these figures. The glossary of terms could likewise be reduced to the first few items since most definitions related to seizure performance assessment are never used in the text.

I fixed a few typos.

Reviewer A:

This a very interesting and well-written article. Very good work, congratulations.

My only general comment is that it might be good to have a 'state-of-the-art' section. This could include if and in what context seizure detection algorithms are currently used. What is their reliability? (Taking into account all the difficulties in the assessment of the performance that are discussed in the text.)

I have a number of very minor on the rest of the text. I list these below in the order of appearance in the text.

The embedded links #Fi do not work (at least not in my browser).

Seizure detection, Second paragraph 'Important additional objectives' can simply be 'Important objectives'.

The last expression in the bracket 'information that is unavailable...' is somewhat difficult to understand. Later on, it is written much better. 'only signal information available up to that point in time'

Seizure detection, third paragraph, last sentence (before 'Motivation and application'): This sentence is quite long and perhaps difficult to read for the general audience.

Motivation and Application, first paragraph '...associated with epilepsy and this uncertainty...' should be '...associated with epilepsy, and this uncertainty...'

The expression 'of the dynamical disease of epilepsy' sounds strange to me (the second 'of').

Just before 'Algorithms for seizure detection', you talk about that the closed-loop therapy could be applied when and where needed. Is it really so that one could spatially target the therapy? In particular, would this be in different brain regions for different seizures of the same patient?

'Algorithms for seizure detection', towards the end. It is written 'Statistical differences between the measures ... are compared...' Are the differences compared? Or is it that rather the measures are compared, and it is tested whether the differences are significant?

'What is a seizure and when to detect it', beginning At the passage 'Unfortunately, these questions', the reference of 'these' is not so clear anymore, since some text is placed between it.

The last sentence of this paragraph 'There is an ongoing...' is not really connected to the rest of the text. The following paragraph is somewhat unclear. It contains two very long sentences, which might be difficult to decipher for the non-expert reader.

'Minimizing computational cost' I was surprised to read that. Why are only such limited processors approved? Is it because of the heat faster processors produce? It would be good if this could be clarified.

Figures: 1, caption. There is a 'ration' which should be 'ratio'. 2. This figure misses y-labels. 3. The dashed and dotted curves in the lower panel are difficult to distinguish

Responses to Reviews

Thank you to the reviewers for the many important and helpful comments and careful reviews of this article. In this section I will place each item from the reviews (in italics) along with my respective responses (in bold).

Response to Reviewer A Comments

This a very interesting and well-written article. Very good work, congratulations. Thank you.

My only general comment is that it might be good to have a 'state-of-the-art' section. This could include if and in what context seizure detection algorithms are currently used. What is their reliability? (Taking into account all the difficulties in the assessment of the performance that are discussed in the text.) While I think this is an important comment, I have not implemented the suggestion. While I agree with you that it would be ideal to have a “state of the art” section, and perhaps should thank you for the opportunity to publicize my work on the subject (since I think the SDA architecture our group’s work provides may be the best available approach to seizure detection, at least for intracranial recordings), I do not think it is possible to write such a section at present. The reason is that there have been no prospective comparisons of multiple methods on any reasonably sized, common data set. Comparing FP rates and sensitivity/specificity results from one article to the next gives some indication, but does not answer the question of which SDA is the better between available options. For this reason, I have started a list of “commercially available SDAs” since these are at least available, whereas many published algorithms can’t even be tested because their implementation details aren’t always accessible.

I have a number of very minor on the rest of the text. I list these below in the order of appearance in the text. The embedded links #Fi do not work (at least not in my browser). I have updated the links to fix this problem.

Seizure detection, Second paragraph 'Important additional objectives' can simply be 'Important objectives'. Done.

The last expression in the bracket 'information that is unavailable...' is somewhat difficult to understand. Later on, it is written much better. 'only signal information available up to that point in time' I changed it to read “only using information that is available at the time the detection is made.”

Seizure detection, third paragraph, last sentence (before 'Motivation and application'): This sentence is quite long and perhaps difficult to read for the general audience. I changed this paragraph to improve clarity. See my response to Reviewer B's related comment below.

Motivation and Application, first paragraph '...associated with epilepsy and this uncertainty...' should be '...associated with epilepsy, and this uncertainty...' Done.

The expression 'of the dynamical disease of epilepsy' sounds strange to me (the second 'of'). Agree. I changed it to “and to improve understanding of epilepsy as a dynamical disease.”

Just before 'Algorithms for seizure detection', you talk about that the closed-loop therapy could be applied when and where needed. Is it really so that one could spatially target the therapy? In particular, would this be in different brain regions for different seizures of the same patient? Yes, therapies really can be spatially targeted. Simple examples include stimulating left or right temporal lobe depending upon whether the seizure begins on the left or right, etc. I revised the last few sentences of this section, which discussed closed-loop therapy to try to state ideas more clearly.

'Algorithms for seizure detection', towards the end. It is written 'Statistical differences between the measures ... are compared...' Are the differences compared? Or is it that rather the measures are compared, and it is tested whether the differences are significant? Your point is a good one. I modified the sentence to state: “Differences between measures in the most recent moving window(s) are typically compared to reference or background values to identify statistically significant changes associated with the seizure activity.”

'What is a seizure and when to detect it', beginning At the passage 'Unfortunately, these questions', the reference of 'these' is not so clear anymore, since some text is placed between it. I agree. I modified the “unfortunately” sentence to make it clearer. It now reads: “Unfortunately, answers to the "what is a seizure?" and "when to detect it?" questions remain elusive since there is currently no consensus or objective definition of what constitutes a seizure.”

The last sentence of this paragraph 'There is an ongoing...' is not really connected to the rest of the text. The following paragraph is somewhat unclear. It contains two very long sentences, which might be difficult to decipher for the non-expert reader. I agree. I reordered and restructured the entire section to improve clarity.

'Minimizing computational cost' I was surprised to read that. Why are only such limited processors approved? Is it because of the heat faster processors produce? It would be good if this could be clarified. Reasons for this include the cost of gaining regulatory approval for state of the art processors and the manufacturing constraints imposed by regulatory agencies. I included this statement and cited a lecture on the subject given at the Freiburg Seizure Prediction Workshop by Jon Werder of Medtronic.

Figures: 1, caption. There is a 'ration' which should be 'ratio'. 2. This figure misses y-labels. 3. The dashed and dotted curves in the lower panel are difficult to distinguish The SDA output ratio is dimensionless. I included this note in the caption. I also revised Figure 3 so the curves are all solid but used different colors to make them much easier to distinguish.

Response to Reviewer B Comments

This is a comprehensive article about the problems and pitfalls associated with an automated algorithmic detection of epileptic seizures in electrophysiological time series, written by one of the experts and pioneers in the field. The section on Challenges in Seizure Detection is particularly helpful. Thank you.

I only have a few minor suggestions for improvement. 1st paragraph: A clear distinction between the fields of seizure detection and seizure prediction would be helpful. I agree. I have revised this paragraph to calrify the distinction.

Fig. 1 and 2: Axis labels area barely legible. I have enlarged the size of the figures on the page, but the user may still need to click on the figure and zoom in to see all the details.

Fig. 2: Do all of the algorithmic detections correspond to actual seizures? If not, which of them are true positives and which are false positives? All the detections shown in Figure 2 were true positives, corresponding to events that were scored as seizures. I clarified this in the figure legend.

Fig. 5: This figure should have double width so the scaling on the abscissa remains constant when the second half of the data is added. Fig. 8: Same as Fig. 5. I have modified Fig. 5 and what was Fig. 8 so the scaling on the abscissa remains constant when the second half of the data is added. I have also moved the example that comprised Figs 6-8 in the previous draft to the very end of the page (now Figs. 9-11) and am willing to remove Figs. 9-11 altogether if you think the inclusion of a second example here is excessive given the overall length of the article (see your next comment and my response below).

The paragraph on SDA Performance Assessment is short and succinct, which is good given the overall length of the article. I am not sure, however, whether Figs. 10 and 11 can be easily understood without a lot more explanation. So I wonder if it might be more helpful to just omit these figures. The glossary of terms could likewise be reduced to the first few items since most definitions related to seizure performance assessment are never used in the text. I think the examples of what were Figs. 10 and 11 (now Figs. 7-8) convey important ideas that should be included for educating readers about SDA assessment. I have revised the captions to try to make them more easily understood. If you can identify ambiguities or other things that need further explanation, I would prefer to rectify this without removing the examples. I shortened the glossary of terms to remove the performance assessment terms not mentioned in the text.

I fixed a few typos. Thank you.

Response to Previous Anonymous Recommendation

RECOMMENDATION : The author is recommended to consider probabilistic approaches to early detection of Seizures. These techniques fuse the multivariate data using model of normality by modelling the underlying data density. A probabilistic threshold is obtained from the model of normality to detect the seizure events as deviations away from normality. References [1] Stephen Roberts, Lionel Tarassenko, Probabilistic Resource Allocating Network for Novelty Detection, Neural Computation, March 1994, Vol 6, No 2, Pages 270-284. [2] P.E. McSharry, T. He, L.A. Smith and L. Tarassenko, "Linear and non-linear methods for automatic seizure detection in scalp electroencephalogram recordings", Journal of Medical and Biological Engineering and Computing, vol. 40, no. 4, pp. 447-461, July 2002.

I do not agree with this reviewer that the references mentioned are important enough to the field to warrant mention in this article. The “probabilistic approach” they mention is incorporated in some sense in most methods used for seizure detection (e.g., in threshold determination), including methods predating these articles. At the same time, Gaussian models are very basic approximations. The McSharry et al. article is a good one, but I do not see a specific novel or important contribution to the field sufficient to warrant its inclusion here.