This study deals with a long lasting problem of automated detection of interictal epileptiform discharge (IED) in intracranial and scalp EEG using deep learning.
The study covers the recent literature adequately.
However, the methods need further clarifications. The datasets are rather small (7 and 5 patients) and should be described in more detail. For example the behavioral state of patients during recording (sleep, rest with closed eyes?) is not specified. This can have potentially impact on the results. What was the length of the selected segments? In dataset 2 only the most active channel was selected, that means that only 5 channels were used from the second dataset? Were the data in the second dataset marked by both reviewers and only the spikes were the reviewers agreed were taken into account? This is unclear. I would also suggest comparing the results of this work with the detector designed by Janca et al. Since it is the same dataset the comparison should be very easy. I also suggest the analysis of impact of the pathology on the results (if the volume of the data is sufficient). In dataset 1 5 out of 7 patients suffered from FCD which could have had impact on the results.
In the abstract the authors claim that their method outperforms SVM and random forest (RF) classifiers. This is not supported by the results. The authors provide statistical comparison between SVM and RF but they do not do the same for RF and IEDnet. Judging by Figure 4 the improvement of RF was only marginal. Why was not the same comparison done for dataset 2 where IEDnet yielded worse results? Some parts of the results section should be moved to the methods section - such as the utilization of DWT to quantify the distances, description of the impact of training sample size and sampling rate, robustness to noise, etc.