每一篇论文都是科研人员的心血，投稿后“漂泊”在外（under review），“老母亲”在家日夜惦念(wait…)。终有一天，它带着沉甸甸的“包裹”回来了 (reviewer comments)，至于“包裹”里是惊喜还是惊吓，包藏着一段怎样艰辛的历程，我们又该如何对其“梳妆打扮”甚至“改头换面” (response to reviewers’ comments)，送他“安全”返回 (revised submission) 并抵达终点 (accept) 呢……小编带你且看且分析。

审稿人的comments千差万别，不过依旧可以找到些许规律可循。小编认为，回复审稿人主要需要把握四点：端正个人态度、吃透反馈问题、分析回复思路、整合回复语言。

Tip 1 端正个人态度

一般情况下，审稿人多为高校或研究机构的科研工作者，他们或常年奋斗在科研一线，经验丰富；或早年“叱咤”学术圈，功成身退。不管是哪种审稿人，他们也都是人，不是神。因此，存在对作者论文的内容理解不到位，所提出的问题不是非常合情合理的可能，甚至可能存在观点偏颇；亦或审稿人对作者论文所述观点理解极深，所提出的问题异常尖锐，搞的作者一时之间不知如何回复。但无论什么样的审稿人，提出什么样的问题，作者首先在态度上应该做到不卑不亢、张弛有度。带着学术交流与沟通学习的目的来审视这些问题，卸下心中的惶恐与不安，保持自信。良好的开端是成功的一半，接下来才是重头戏。

Tip 2 吃透反馈问题

认真逐字逐句研读审稿人的comments，这是做好回复的重要基础。审稿人的问题基本上可以分为参考文献问题、论文内容/写作问题、论文方法问题三类。接下来我们通过举例逐个分析。（例子涉及专业知识，请勿纠结）

1、参考文献问题

具体案例：

(1) Reviewer#1, Concern # 1:

Add more refences to the literature when you talk about the pre-processing, feature extraction, machine learning, and deep learning methods. I suggest the following references.

① A.Alqudah, Ali Mohammad. “An enhanced method for real-time modelling of cardiac related

biosignals using Gaussian mixtures.” Journal of medical engineering & technology 41.8 (2017):

600-611.

② B.Alqudah, Ali Mohammad, et al. “Developing of robust and high accurate ECG beat

classification by combining Gaussian mixtures and wavelets features.” Australasian physical &

engineering sciences in medicine 42.1 (2019): 149-157.

③ C.Alquran, H., et al. “ECG classification using higher order spectral estimation and deep learning techniques.” Neural Network World 29.4 (2019): 207-219.

(2) Reviewer#2, Concern # 2:

Literature improvement is still needed. Please compare your work with some recent work, e.g.,

① J. Huang, B. Chen, B. Yao and W. He, “ECG Arrhythmia Classification Using STFT-Based Spectrogram and Convolutional Neural Network,” in IEEE Access, vol. 7, pp. 92871-92880, 2019.

② B. Hou, J. Yang, P. Wang and R. Yan, “LSTM-Based Auto-Encoder Model for ECG Arrhythmias Classification,” in IEEE Transactions on Instrumentation and Measurement, vol. 69, no. 4, pp. 1232-1240, April 2020.

In Sec. II.C, when “image classification, object detection, and image segmentation” are mentioned, please also cite some recent work like:

③ T. D. Pham, K. Wardell, A. Eklund and G. Salerud, “Classification of short time series in early Parkinsonʼ s disease with deep learning of fuzzy recurrence plots,” in IEEE/CAA Journal of Automatica Sinica, vol. 6, no. 6, pp. 1306-1317, November 2019.

2019. Y. Tian, X. Li, K. Wang and F. Wang, “Training and testing object detectors with virtual images,” in IEEE/CAA Journal of Automatica Sinica, vol. 5, no. 2, pp. 539-546, Mar. 2018.

案例剖析：

上述审稿人的2个问题均涉及参考文献问题，主要聚焦整篇文章的参考文献的质量与选取，或者针对某一部分内容，如（2）中审稿人提出预处理、特征提取、机器学习和深度学习方法部分的参考文献需要进一步优化。同时，所有审稿人均列出了可供引用的参考文献。

针对于上述涉及参考文献类问题的审稿意见，我们首先要对审稿人推荐的参考文献进行筛选，对于与文章内容没有直接关系的参考文献不予引用（但要说明原因），对于与文章内容有密切关系的参考文献则“能引尽引”（不可否认的是，审稿人推荐的文献很可能是本人实验室发表的与审稿人研究领域相关的文献，旨在增加其科研成果的曝光度）。与此同时，作者也可根据审稿人的意见对参考文献进行合适的增删，涉及性能比对等内容时，最好选择近两年的参考文献，以保证论文的“优势”。

回复策略：

三步走策略：首先是感谢审稿人的审稿工作，然后逐个分析审稿人推荐的参考文献，并适当与本文的研究内容进行比较，保留需要参考的文献，去除与本文无关的参考文献并说明原因，最后阐明文中的修改位置和内容。据此，我们对上述问题（1）进行了回复以供参考。

(1) Reviewer#2, Concern # 2:

Author response:

We want to thank reviewer for constructive and insightful criticism and advice about the reference. We addressed all the points raised by the reviewer as summarized below.（表示感谢）

The reference ① above proposed an ECG arrhythmia classification method using two-dimensional (2D) deep convolutional neural network (CNN). Using ECG recordings from the MIT-BIH arrhythmia database as the training and testing data, the classification results show that the proposed 2D-CNN model can reach an averaged accuracy of 99.00%. Compare with our work, even though the performance metrics (accuracy, sensitivity, and specificity) of our paper are more comprehensive than the metrics (only accuracy) of ① the average accuracy from reference ① is better than our model result. And the deep learning method of STFT-Based Spectrogram provide a new idea in future work. (分析与比较)

The reference ② above introduced a novel deep learning approach that integrates a long short-term memory (LSTM)-based auto-encoder (AE) network with support vector machine (SVM) for electrocardiogram (ECG) arrhythmias classification. The proposed method achieved average accuracy, sensitivity, and specificity of 99.45%, 98.63%, and 99.66%, respectively, in the beat-based cross-validation approach based on the Advancement of Medical Instrumentation (AAMI) standards. The performance is superior to our model result (average accuracy, sensitivity, and specificity of 98.00%, 96.17%, and 96.38%). We think the LSTM-based auto-encoder (AE) network is a positive strategy, which can effectively extract the characteristic information of time series signals. (分析与比较)

Hence, we supplement some recent work in Table 10. The comparative literature is comprehensive, which not only include literature of the machine learning method and deep learning method, but also the main work is the two or multi-classification of arrhythmia signal. And most literature also is recent years (2019 and 2020). (修改内容)

Meanwhile, we also add the recent literature like ③ and ④ above in Sec. II.C. when “image classification, object detection, and image segmentation” are mentioned. The classic literature and latest literature complement each other to make the reference literature more substantial and reasonable. (修改内容)

2、论文内容/写作问题

具体案例：

(1) Reviewer#3, Concern # 2:

In Section C. Machining Learning Methods, the authors mention many disadvantages of the reviewed methods, such as poor robustness, overfitting and bad performance in practical applications. Since many shortcomings have been mentioned, the authors are supposed to find a way to solve them. However, in the experimental part, it is not clear whether the proposed method can overcome the shortcomings mentioned in Page 4 (主要问题1). Only results with high performance metrics are presented. It makes the focus of this paper unclear (主要问题2). In my view, these shortcomings mentioned in Page 4 are common among existing machine learning methods. The authors are encouraged to dig deeper and conclude the disadvantages and challenges of existing methods for ECG signal detection and classification, and then propose your method and highlight its difference and superiority (修改意见1). In my view, the major contribution of this paper is the new design of a network architecture. The authors are encouraged to highlight this and compare with other related and similar networks (修改意见2).

(2) Reviewer#1, Concern # 1:

I do remember that I have reviewed this paper, and the current version is an improved version of that paper. I have also seen that the authors did a good job in improving the quality of paper, although I personally do not like the organization and format of the paper.

It also seems that the proposed method could not outperform the best results from the literature (主要问题1). But in my opinion, it is still worthwhile publishing this paper to let others know that this approach has already been tested and compared to other methods could achieve a relatively less accurate result.

Finally, I suggest that the authors improve the quality of their paper and avoid English problems (主要问题2), and if possible, shorten the paper, as it is quite lengthy (主要问题3).

案例剖析：

上述审稿人的两个问题均涉及论文内容/写作问题，而且两个问题比较长，所以要分析提取出主要问题。问题（1）中涉及两个主要问题，主要问题（1）是实验部分没有清楚表明论文所提模型解决了Section C中文献综述所阐述的目前方法的缺陷；主要问题（2）是论文模型仅展示了非常好的性能，但是论文的焦点不是很清晰。这两个主要问题其实表述较为模糊，很难找到很好的切入口进行回复和并作出针对性的修改。但是，审稿人同时也给出两点修改意见。所以，针对这类问题，我们只需根据审稿人给出的两个修改方向进行修改即可。问题（2）中涉及了3个主要问题：论文模型性能并没有优于对比文献的所有参考文献（虽然审稿人表明模型性能不是最优，但论文仍有价值。我们依然要对这个问题进行回复）、论文英语语法需要修正和论文内容需删减。

回复策略：

依然三步走策略：首先是感谢审稿人的审稿工作，然后分析审稿人comments中的主要问题和修正意见，最后针对审稿人所提出的主要问题进行逐条回复并修正。

具体回复：

(1) Reviewer#3, Concern # 2:

Author response: 

Thank you very much for your careful review and constructive suggestions with regard to our manuscript. Meanwhile, I am also very grateful to you for the specific modification strategy. (表示感谢) I have made comprehensive and detailed changes according to your suggestions. On one hand, we dig deeper and conclude the disadvantages and challenges of existing methods in section II. On the other hand, we also highlight the proposed model’s difference and superiority and compare with other related and similar networks. Below is our point-by-point modification to the referee’s suggestions.

(1) The disadvantages and challenges of existing methods for ECG signal detection and classification. (问题剖析)

Broadly speaking, the fundamental disadvantages and challenges of existing machine learning methods for ECG signal detection and classification are that hand-crafted extracted feature not only greatly affects the accuracy of the algorithm, but also consumes a lot of calculation time and cost. The deep convolutional neural network is essentially realized by stacking automatic encoders. Considerable feature representational power effectively reveals unknown abstract features of input signals. It can achieve self-learning through end-to-end model design. Meanwhile, the radical problem of both methods is that they only focus on how to propose a better model, but do not pay attention to data processing issues: such as data denoising, data augmentation, and multi-scale data training and testing. The data preprocessing of signals should be focus on because signals and images are different data types.

Hence, in this work, a more accurate and robust method based on deep learning is proposed to identify five different types of arrhythmia signals. The proposed model not only pays attention to the superiority of model design but also presents the importance of data processing in this paper. The final results also prove that the application of ECG signal classification using the convolutional neural network is reliable. The deep learning architecture outperforms the hand-crafted feature extractors assembled by machine learning models in terms of classification accuracy, sensitivity, specificity, and confusion matrix.

(2) The contributions of this work and model’s superiority are as follows: (问题剖析)

① We propose an end-to-end plain-CNN architecture and two MSF-CNN architectures (A and B) to replace additional hand-crafted feature extraction, selection, and classification using machine learning methods. The plain-CNN is a baseline model, the MSF-CNN A and B are implemented on this baseline. Thus, it significantly enhances the performance against recent state-of-the-art studies. The best model MSF-CNN architecture B achieves an average accuracy, sensitivity, and specificity of 98.00%, 96.17%, and 96.38%, respectively. This illustrates the method with residual learning and group convolution blocks has a profound effect on the feature learning of the model. The results of ablation experiments show that our proposed biometric recognition and diagnosis network with residual learning (MSF-CNN B) achieves a rapid and reliable diagnosis approach on ECG signal classification, which has the potential for introduction into clinical practice as an excellent tool for aiding cardiologists in reading ECG heartbeat signals. (分层回答)

② Moreover, the signal processing problems are fully considered. We first design multi-scale input signals, including 251 samples (named set A) and 361 samples (named set B). This design can improve the generalization ability of the model by extracting multi-scale signal features. Then, the signal denoising and data augmentation also are implemented in this paper. (分层回答)

③ In addition, we present six sets of detailed ablation experiments on ECG signal classification and achieve excellent performance metrics. Our residual learning network with group convolution (including parallel and concatenation group convolutional block) can easily achieve accuracy gains from greatly increased depth, producing results substantially better than the plain-CNN and MSF-CNN A. It also demonstrates deep learning can be leveraged as feature learning mechanisms to classify biomedical signals. (分层回答)

④ In particular, we compare the results from our model to recent state-of-the-art methods. Additionally, detailed analysis and comparison are presented in this paper. (分层回答)

Finally, in this manuscript, we replace some old references with the recent methods published in 2019 and 2020 in Table 10. Meanwhile, we have also done a full comparison and analysis.

(2) Reviewer#1, Concern # 1:

Author response: 

We would like to thank the reviewer for their thoughtful review of our manuscript (表示感谢). According to the reviewer’s comments, I think that the reviewer has no serious problems about this manuscript. There are three problems that require minor edits. The first minor problem is that the proposed method could not outperform the best results from the literature. The second minor problem is that the English expression needs to be improve. And the last minor problem is that the paper needs to be shorten if possible.

(1) The first minor problem is that the proposed method could not outperform the best results from the literature. (问题剖析)

In [62], the performance is superior to our models’ result (average accuracy, sensitivity, and specificity of 98.00%, 96.17%, and 96.38%). We think the LSTM-based auto-encoder (AE) network [62] is a positive strategy, which can effectively extract the characteristic information of time series signals. Most importantly, the method in [62] is the key point worthy of our study in future work. We will fully consider the optimization methods of [62] in our future work.

Compared with [62], our research mainly has the following differences:

First, compared with the LSTM-based auto-encoder network in [62], our model is more lightweight and less computationally expensive. The LSTM is a bidirectional model, which is utilized to extract the bidirectional information from the forward model (operating from t₀, t₁, …, to t_n.) and backward model (operating from t_n, t_n-1, …, to t₀.) at the same time. There is no doubt that the advantage will also cost a lot of computational expensive. (分层回答)

Second, in recent years, the CNN (Convolutional Neural Network) and RNN (Recurrent Neural Network) is two popular deep learning methods to process the time series data. Both methods have their own advantages. The LSTM is a recurrent neural network that has an input layer, two hidden layers, and an output layer. The hidden layers are a memory block included three gate mechanisms called input gate, forget gate, and output gate. These gate mechanisms control the amount of information and the extraction of features. (分层回答)

In additional, [62] is also the latest reference published in 2020 that I added in the last revision. The reason why I cite this paper with better performance index is we think the auto-encode network based on LSTM is a positive idea. Our paper has gone through a long review process, during which the associate editor and reviewer have been changed. Compared with [62], the performance advantage is really not obvious. Hence, we will try to propose a lightweight model based on LSTM to deal with this problem in future work. (分层回答)

(2) The second minor problem is that the English expression needs to be improve. (问题剖析)

We tried our best to improve the expression of grammar, typos and professional representation of this manuscript and made many changes in the manuscript. And we also leverage 3rd party service for language polishing. we take great care to revise the grammars, typos and professional representation of our manuscript.

(3) And the last minor problem is that the paper needs to be shorten if possible. (问题剖析)

This paper has been reviewed by many reviewers, and a lot of contents have been added to the original text according to the requirements of the reviewers. Therefore, many revisions of this paper lead to the final paper being too long. In this manuscript, we also deleted some content appropriately as soon as possible. In section II. Related work, we deleted the detailed description about the references’ contents on the basis of retaining relevant research references. In section III and VI, we also deleted inessential contents. （修改内容）

3、论文方法问题

具体案例：

(1) AE’s comment, Concern # 1:

Also, finding optimal parameters to ensure the network performance is important by using advanced intelligent optimization methods, e.g.,

① J. J. Wang and T. Kumbasar, “Parameter optimization of interval Type-2 fuzzy neural networks based on PSO and BBBC methods,” IEEE/CAA J. Autom. Sinica, vol. 6, no. 1, pp. 247−257, Jan. 2019 and S. Gao, et al.,

② “Dendritic neuron model with effective learning algorithms for classification, approximation and prediction,” IEEE Transactions on Neural Networks and Learning Systems, 30(2), pp. 601 - 614, Feb. 2019. I would suggest the authors to treat it as future work.

(2) Reviewer#2, Concern # 3:

In [63] and few others, the accuracy is better in others’ work. The authors need to discuss where their work stands with other works on 5-class classification problem.

案例剖析：

上述问题（1）中涉及模型参数优化问题，并给出具体的模型参数优化的参考文献。面对这样的问题往往很迷茫，如果重新选择参数优化方法，那就面临要再次做具体的实验，并且和本文的实验进行分析比较。小编感觉这样修正内容过多，而且可能会偏离本论文所论述的具体问题，很没有必要再增加实验，其实审稿人提出的这个问题小编个人感觉重点不是很突出，推荐将这一问题转换成参考文献的问题。首先依然是阐述并筛选出有效的参考文献，然后分析参考文献的方法与本文的方法的异同点，重点突出本文的优势和审稿人推荐文献所用方法的可取之处。

上述问题（2）也非常具有代表性，主要表述作者提出的参考文献[63]的性能要优于文章模型的性能，可谓是作者自己给自己挖了个大坑。但是，作者原本意图是想阐述文献[63]中有些优势可以借鉴，是以后研究的重要方向，客观的来讲，每篇文章不可能尽善尽美，除了优势之外，对于limitation的阐述也是非常重要的。面对这一问题，主要策略首先还是作者需要阐明本意，然后分析比较两篇文章，突出本文优势。

回复策略：

重要的“三步走策略”说三遍：首先是感谢审稿人的审稿工作，然后深入剖析审稿人comments中的主要问题，不用遗漏审稿人提出的所有细节问题，最后针对审稿人所提出的主要问题进行逐条回复并修正，重点要突出本论文方法设计的主要优势。

具体回复：

(1) AE’s comment, Concern # 1:

Author response: 

Thanks a lot for the AE’s comments about the parameter optimization. In this paper, we adopt some effective parameter optimization methods. （表示感谢）

First, back-propagation algorithm. In this paper, we proposed three convolutional neural networks. The parameters update of the models mainly rely on the back-propagation algorithm. The gradient is backed to the shallow layers of network so that implement the update of weight and bias. (分层回答)

In addition, in the experiment, the L2 norm of the model parameters (equation (1)) is implemented to optimize gradient and relieve overfitting. Specifically, the threshold is set to 0.5 to stabilize the training process.

              

where  is the loss function with L2 regularization and  is the cross-entropy loss function from equation (9). denotes a penalty factor, which is to balance the goal of achieving better training results and keeping smaller parameter values. Thus, the regularization can avoid overfitting effectively by narrowing down all the parameters.  describes the weight of  layers. (分层回答)

As mentioned in reference ① above, in this paper, interval type-2 fuzzy neural networks (IT2FNNs) is the hybridization of interval type-2 fuzzy systems (IT2FSs) and neural networks (NNs). Big bang-big crunch (BBBC) optimization and particle swarm optimization (PSO) are applied in the parameter optimization for Takagi-Sugeno-Kang (TSK) type IT2 FNNs. The employment of the BBBC and PSO strategies can eliminate the need of backpropagation computation. The computing problem is converted to a simple feed forward IT2 FNNs learning. The adoption of the BBBC or the PSO will not only simplify the design of the IT2 FNNs but will also increase identification accuracy when compared with present methods. (分析与比较)

In reference ② above, in this paper, these problems, such as  the hard understanding problem, the slow and difficult training problems, and the difficulty to scale them up, motivate us to develop a new dendritic neuron model (DNM) by considering the nonlinearity of synapses, not only for a better understanding of a biological neuronal system, but also for providing a more useful method for solving practical problems. To achieve its better performance for solving problems, six learning algorithms including biogeography-based optimization, particle swarm optimization, genetic algorithm, ant colony optimization, evolutionary strategy, and population-based incremental learning are for the first time used to train it. The best combination of its user-defined parameters has been systemically investigated by using the Taguchi’s experimental design method. (分析与比较)

In the future, we will implement the parameter optimization methods mentioned above to achieve its better performance.

At last, in this manuscript, we found many grammatical and representational errors in the paper. Hence, we invite a 3rd party service for language polishing before resubmitting. The grammar and expression have been greatly improved in this paper. We would like to submit the enclosed manuscript, which we wish to be considered for publication in IEEE Access. I carefully revised all the questions given by two reviewers. Meanwhile, we have made detailed response in the supplemental files. (分层回答)

(2) Reviewer#2, Concern # 3:

Author response: 

We would like to thank the reviewer for your thoughtful review of our manuscript. Indeed, in [62] (The [63] in last manuscript is renamed as [62] in this manuscript), the performance (average accuracy, sensitivity, and specificity of 99.45%, 98.63%, and 99.66%) is superior to our models’ result (average accuracy, sensitivity, and specificity of 98.00%, 96.17%, and 96.38%). We think the LSTM-based auto-encoder (AE) network used in [62] is a positive strategy, which can effectively extract the characteristic information of time series signals. We will fully consider the optimization methods of this paper [62] in our future work.

Compared with [62], our research mainly has the following differences:

First, compared with the LSTM-based auto-encoder network in [62], our model is more lightweight and less computationally expensive. The LSTM is a bidirectional model, which is utilized to extract the bidirectional information from the forward model (operating from t₀, t₁, …, to t_n.) and backward model (operating from t_n, t_n-1, …, to t₀.) at the same time. There is no doubt that the advantage will also cost a lot of computational expensive. (分层回答)

Second, in recent years, the CNN (Convolutional Neural Network) and RNN (Recurrent Neural Network, LSTM is a replacement of the traditional RNN) is two popular deep learning methods to process the time series data. Both methods have their own advantages. The LSTM is a recurrent neural network that has an input layer, two hidden layers, and an output layer. The hidden layers are a memory block included three gate mechanisms called input gate, forget gate, and output gate. These gate mechanisms control the amount of information and the extraction of features. (分层回答)

Third, it also seems that the proposed method could not outperform the best results from the literature. But in my opinion, compared with most of performance results in Table [9], the results of our paper is advantageous, and it is still worthwhile publishing this paper to let others know that this approach has already been tested and compared to other methods could achieve a relatively less accurate result. Most importantly, the method in [62] is the key point worthy of our study in future work. (分层回答)

In additional, [62] is also the latest reference published in 2020 that I added in the last revision. The reason why I cite this paper with better performance index is we think the auto-encode network based on LSTM is a positive idea. Our paper has gone through a long review process, during which the associate editor and reviewer have been changed. Compared with [62], the performance advantage is really not obvious. Hence, we will try to propose a lightweight model based on LSTM to deal with this problem in future work. (分层回答)

Tip 3 分析回复思路

小编认为“思路”是回复审稿人问题的重中之重，主线明确、思路清晰，能够使得审稿人读起来清楚明了，这一点十分重要。看过上面Tip2中对于具体问题的剖析，相信你已经对小编推荐的回复思路有了大概的了解。首先，无论审稿人给出的问题尖锐与否、合理与否、正确与否，我们都要对其表达谢意，这是对审稿人工作的必要性尊重；其次，分析审稿人的comments，提取所有涉及到的问题，并分段、分层将问题进行描述；第三，针对每个问题进行回复，如果问题需要大量重点阐述，那就需要作者分层、分点、递推式的阐述；最后需要适当描述对于审稿人comments的具体修改方式和修改位置（如第几页第几行）。

Tip 4整合回复语言

对于审稿人的comments逐条回复，突出回复重点、要点以及希望引起审稿人注意的点是整合回复语言的主要工作。另外，回复问题之前，对审稿人的审稿工作表示感谢也是不多余的，毕竟给审稿人留个懂礼貌知礼节的好印象或许有助于审稿人“开开心心”工作。因此，小编通常推荐以下回复模式：

(1) Reviewer#1, Concern # 2: 列出每位审稿人的每条comments

In Table 10, the authors compare their work with others’ work. But do others’ work have same experiment conditions, e.g., dataset and classes?

(2) Author response: 对审稿人的comments进行个人陈述

We want to thank reviewer for constructive and insightful criticism and advice about the Table 9 (Table 4 is deleted in this manuscript, and the Table 10 is named as Table 9) (表示感谢). We addressed all the points raised by the reviewer as summarized below.（分点阐述）

First, we readjusted the contents of Table 9 and supplemented the column of database and main work. In Table 9, we listed the 12 related papers including publication time, main work, database, main approach, and performances (accuracy, sensitivity, specificity). These papers were published in 2019 and 2020 and represented the latest research results. And the main work is the classification of arrhythmia signals. It's just that the number of categories in some papers is different. The used database mainly includes MIT-BIH arrhythmia database and Chinese Cardiovascular Disease Database, of which most studies use MIT-BIH arrhythmia database. This is consistent with our research paper. The main methods used in this comparative literature are machine learning and deep learning, which is mainly to compare the final performance of the two methods.

Second, we also fully analyzed the similarities and differences between the research from others’ work listed in Table 9 and our research in this paper in section V. C. (performance comparison and discussion). The main research work of these references listed in Table 9 is to detect and classify arrhythmia signals. Almost all studies use the MIT-BIH arrhythmia database except reference [53], [56] and [60]. However, the different databases have little impact on the final analysis. Because the original data from two databases need to pre-processing, such as wave detection, normalization, signal segmentation, and signal annotation (This work was completed by experts of the database publishers in the early stage).

In additional, we mainly summarized the research literature on the use of machine learning and deep learning methods, which is to analyze the advantages of deep learning method.

(3) Author action: 简要说明具体的修改方法与具体位置

We firstly updated the manuscript by readjusting the contents in Table 9 and supplementing the column of database and main work in Table 9. In additional, we also fully analyzed the similarities and differences between the research from others’ work listed in Table 9 and our research in section V. C. All the corrections were highlighted using a yellow highlight tool within the new pdf file.

因此，小编推荐的回复模式主要由以下三个部分构成：罗列审稿人的comments（必要时可将其分条描述）；在回复时首先表示感谢，再通过分段、分点、分层的方式阐述观点、修改的内容以及相应的结论；对修改的方法和位置进行简要描述。此外，可以通过改变字体颜色、粗细程度来表示重点信息。

经过一系列的“猛虎操作”，相信你已经可以基本了解审稿意见的回复过程啦。不过具体情况具体分析，不同的研究领域其学术习惯也不尽相同，因此小编郑重提示：以上交流仅供参考，希望大家多多批评指正！