A Hybrid DenseNet-LSTM Model for Epileptic Seizure Prediction

Round 1

Reviewer 1 Report

In this manuscript, the authors establish a novel method of seizure prediction through the application of two deep learning methods, Dense Convolutional Network (DenseNet) and the Long Short Term Memory (LSTM) network. Predicting seizures is of immense practical importance both to patients and to medical device manufactures. Through the application of DenseNet and LSTM the authors show impressive pre-ictal signal classification and an improvement over other methods. Overall, this is an interesting study although I do have some comments (in order of appearance): 

Line 3: should read "patients with epilepsy"

Ln 12: which pre-ictal time window corresponds to the accuracy, etc. reported here? Is this an average? 

Ln 25-26: presumably you mean "neurons" and not nerves? 

Fig 1- All of the figures need more descriptive legends. For this figure, what is the time domain displayed? I'm assuming that this waveform is from a single channel; what channel was used? 

Ln 108- presumably there is one image per channel and time segment. What is the frequency bandwidth over which the EEG channels are decomposed? While you are using wavelets, they can be converted to frequencies. What was the file format for your images? Some formats are lossy.

Ln 124-129, Fig 3: This section was a little unclear to me. Are you saying that you padded the ictal period with 5 minutes preceding the seizure and then measured the preictal period as an additional period prior to this padded period? What is meant by Line 124 "In the case of Interictal"? Doesn't this refer to pre-ictal data? 

In general, what types of seizures did these patients have? Absence, partial, GTCs? The brain areas and total volume of tissue involved during the seizures would be potentially large variables; does this affect your algorithm?

Ln 108: db4 should be defined here, not at line 140

Ln 152- what is CVPR?

Fig 5- what do the different squares and lines represent?

Ln 163- define ReLU

fig 6: the transition layer should be added or depicted in a separate figure.

Ln 170: how is the compression factor determined?

Equations 1-6: Please define all terms.

Eq 6: should read "tanh"

Ln 211: what does it means for features to be classified as sigmoid? Is this a transitional state?

Ln 224: as a negative control, did you ever try to falsely label some interictal periods as pre-ictal and then do classification between the two types of interictal datasets? 

Table 4- all abbreviations should be defined. Also, what do "precision" and "recall" mean?

Ln 236- with the scale on fig 9, the values for the different pre-ictal periods look different, but there isn't a lot of actual difference between them and the model is pretty sensitive regardless with some slight differences for the different time epochs. Did you do statistics to determine if there were true differences between model performance for the different time periods?

Fig 10- which time epoch was used for your data here?

 

 

Author Response

Please see the attachment

Author Response File: Author Response.docx

Reviewer 2 Report

Authors proposed the DenseNet-LSTM, a hybrid deep learning model, for seizure prediction using the transformed EEG data through DWT. The proposed method resulted in the significantly improved performance compared to other two conventional methods.

However, there has been the several previous studies to combine CNN and RNN networks for the same dataset.  Please, add these publications in your citations and find more if there is more recent works. In addition, please add the comparison results with these models to show the improvement of your work.

1) G. Choi, C. Park, J. Kim, K. Cho, T.-J. Kim, H. Bae, K. Min, K.-Y. Jung, and J. Chong, “A novel multi-scale 3d cnn with deep neural network for epileptic seizure detection,” in 2019 IEEE International Conference on Consumer Electronics (ICCE). IEEE, 2019, pp. 1–2.

2) W. Liang, H. Pei, Q. Cai, and Y. Wang, “Scalp eeg epileptogenic zone recognition and localization based on long-term recurrent convolutional network,” Neurocomputing, 2019.

 

And, there were also several studies using DWT to detect seizure for EEG signals. Please, add the following publications in your citations also.

3) R. Akut, “Wavelet based deep learning approach for epilepsy detection,” Health information science and systems, vol. 7, no. 1, p. 8, 2019.

4) P. Boonyakitanont, A. Lek-uthai, K. Chomtho, and J. Songsiri, “A comparison of deep neural networks for seizure detection in eeg signals,” bioRxiv, p. 702654, 2019.

5) A. M. Karim, O. Karal, and F. C¸ elebi, “A new automatic epilepsy serious detection method by using deep learning based on discrete wavelet transform,” no, vol. 4, pp. 15–18, 2018.

 

In Figure 10 of Results section, authors provided the sensitivity and FPR as the comparison metric with other methods. Please, provide other metrics, such as accuracy, specificity, F1-score as an additional information for comparison.

Author Response

Please see the attachment

Author Response File: Author Response.docx

Round 2

Reviewer 2 Report

Thank for providing all information I requested.