Comparison of Bioelectric Signals and Their Applications in Artificial Intelligence: A Review
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsThis study reviews bioelectrical signals and their integration with artificial intelligence (AI) for clinical and technological applications. It examines signal processing challenges, compares AI methodologies, and highlights advances in medical diagnostics, rehabilitation, and assistive technologies, emphasizing AI’s role in improving signal analysis and interpretation. However, the paper suffers from the limitations listed below, which must be “fully” addressed before its reconsideration:
1- The review presents various AI approaches (e.g., CNN, LSTM, GNN) but lacks a direct performance comparison across models for the same dataset. Did the authors attempt a meta-analysis with statistical measures (e.g., F1-score, AUC-ROC) to assess which AI methods consistently outperform others across different bioelectric signals?
2- The paper mentions deep learning models but does not account for the impact of training data quantity. Since deep learning models require extensive datasets, how do the authors justify their effectiveness for bioelectrical signals when many public datasets contain fewer than 10,000 samples? Have they assessed overfitting risks in AI models due to limited training data?
3- Many bioelectrical signals, such as EEG and ECG, exhibit nonlinear behaviors. Why does the review primarily focus on traditional AI approaches instead of exploring chaos theory, fractal analysis, or nonlinear dynamical system modeling, which could improve signal characterization?
4- The review references various AI applications but does not indicate if the cited studies used the same evaluation criteria. Did the authors conduct a systematic review following PRISMA guidelines, or is the selection of cited works arbitrary? How do they ensure that papers reporting high accuracy do not suffer publication bias?
5- The authors optimized their machine learning model by employing five deep learning architectures (VGG-16, ResNet-50, DenseNet-121, ResNext-50, and ViT) and comparing different mother wavelets for signal conversion into scalograms, using a five-fold cross-validation strategy to enhance accuracy. This is valuable, but please mention that other strong architectures could be leveraged, such as DeepLab (https://doi.org/10.1038/s41467-020-18147-8) and EfficientNet (https://doi.org/10.1038/s41467-024-53993-w). Please briefly introduce these two architectures and reference the referred papers.
6- The paper treats EEG, ECG, and EMG signals as independent categories, yet multimodal AI approaches (e.g., EEG-EMG fusion for neurorehabilitation) could offer superior accuracy. Why did the review not explore hybrid models integrating multiple bioelectrical signals for enhanced classification?
7- The paper does not explore the ethical implications of AI in bioelectrical signal processing. What safeguards should be in place to prevent AI biases from disproportionately affecting certain patient populations (e.g., age, gender, ethnicity-based variations in EEG/ECG morphology)?
Author Response
Comment 1: The review presents various AI approaches (e.g., CNN, LSTM, GNN) but
lacks a direct performance comparison across models for the same dataset.
Did the authors attempt a meta-analysis with statistical measures (e.g., F1-
score, AUC-ROC) to assess which AI methods consistently outperform others
across different bioelectric signals?
Response:
We appreciate your valuable comments on our research. The main objective
of this study is to compare the various applications of bioelectrical signals in
AI. Papers utilizing public datasets have been included in the review, with a
particular focus on those employing diverse methodological approaches. How-
ever, this review does not focus on comparing studies using the same dataset,
as its purpose is to cover as many approaches as possible, considering vari-
ables such as signal extraction methods, data processing strategies, and oth-
ers. In addition, the analysis performed in this study focuses on comparing
results between different works, which implies that the metrics evaluated are
not always uniform. Additionally, the review examines each type of signal
independently, without establishing direct comparisons of model performance
between different types of bioelectrical signals.
Comment 2: The paper mentions deep learning models but does not account for the im-
pact of training data quantity. Since deep learning models require extensive
datasets, how do the authors justify their effectiveness for bioelectrical signals
when many public datasets contain fewer than 10,000 samples? Have they
assessed overfitting risks in AI models due to limited training data?
Response:
Thank you for the reviewer’s insightful observation. In response, it is empha-
sized that the revised manuscript explicitly addresses the issue of data set lim-
itations in several sections. For example, studies such as those by Kulyabin et
1al. and Posada et al. are discussed in detail, showing how synthetic data gen-
eration (via CGAN) and data augmentation techniques were employed to mit-
igate the limitations imposed by small data sets. The revised manuscript also
emphasizes that overfitting remains a concern in studies working with not-
so-large samples, and the authors of these studies suggest expanding datasets
and promoting open data sharing to improve generalizability. These aspects
have been clarified to make the discussion more explicit and relevant in the
context of deep learning applications to bioelectrical signals.
Comment 3: Many bioelectrical signals, such as EEG and ECG, exhibit nonlinear be-
haviors. Why does the review primarily focus on traditional AI approaches
instead of exploring chaos theory, fractal analysis, or nonlinear dynamical
system modeling, which could improve signal characterization?
Response:
The main objective of our research is to present different works and their
applications in the field of artificial intelligence. In addition, it adopts a
comparative approach to analyze the differences in the use of bioelectrical sig-
nals in different contexts. However, we do not intend to perform an analysis
on how to improve bioelectrical signals for further use in the field of AI or
seek a better result than the results obtained by different authors. However,
these points have been clarified in the abstract and introduction of our work.
Comment 4: The review references various AI applications but does not indicate if the
cited studies used the same evaluation criteria. Did the authors conduct a
systematic review following PRISMA guidelines, or is the selection of cited
works arbitrary? How do they ensure that papers reporting high accuracy
do not suffer publication bias?
Response:
Many thanks to the reviewer for this valuable observation. As clarified in the
revised manuscript, the paper follows a narrative literature review approach
rather than a systematic review according to PRISMA guidelines. The aim of
this review is to provide a broad, multidisciplinary overview of the use of AI
in bioelectrical signal analysis, highlighting representative studies of different
AI modalities and techniques. While recognizing the variability of evaluation
metrics among the cited papers, we have attempted to provide clear summaries
in tabular form so that readers can make comparisons where appropriate. We
also added an explanation setting out potential limitations in studies reporting
high accuracy, and suggested corroboration with different reference datasets
or open repositories to improve transparency and comparability.
2Comment 5: The authors optimized their machine learning model by employing five deep
learning architectures (VGG-16, ResNet-50, DenseNet-121, ResNext-50, and
ViT) and comparing different mother wavelets for signal conversion into scalo-
grams, using a five-fold cross-validation strategy to enhance accuracy. This is
valuable, but please mention that other strong architectures could be lever-
aged, such as DeepLab and EfficientNet. Please briefly introduce these two
architectures and reference the referred papers.
Response:
We appreciate the contribution of papers describing the application of these
models. A brief introduction to EfficientNet with Feature Pyramid Networks
(FPN) and DeepLab v3 was presented, highlighting its potential in the analy-
sis of bioelectrical signals. However, although DeepLab v3 has shown promis-
ing results in the healthcare domain, it has not yet been explicitly applied to
bioelectrical signals. On the other hand, EfficientNet has been used in civil
engineering and construction research but not in analyzing these types of sig-
nals. These models were mentioned as an additional comparison alternative
to the previously used by the author [49] cited in the paper.
Comment 6: The paper treats EEG, ECG, and EMG signals as independent categories, yet
multimodal AI approaches (e.g., EEG-EMG fusion for neurorehabilitation)
could offer superior accuracy. Why did the review not explore hybrid models
integrating multiple bioelectrical signals for enhanced classification?
Response:
The reviewer is sincerely thanked for his insightful commentary. It is recog-
nized that multimodal AI models in particular those integrating EEG, EMG
and ECG signals represent a growing and promising area of research. How-
ever, the aim of the present review is to provide a structured and comparative
overview of AI methods applied individually to the main categories of bioelec-
trical signals. By analyzing each modality independently, we aim to highlight
the unique methodological challenges, preprocessing techniques, and learning
architectures associated with each signal type. Although the integration of
multiple modalities is indeed valuable, a detailed exploration of hybrid or fu-
sion models was beyond the scope of this review. This limitation has now
been clarified in the introduction by including a brief comment to highlight
multimodal learning as an important direction for future work.
Comment 7: The paper does not explore the ethical implications of AI in bioelectrical
signal processing. What safeguards should be in place to prevent AI biases
from disproportionately affecting certain patient populations (e.g., age, gen-
der, ethnicity-based variations in EEG/ECG morphology)?
Response:
3We have corrected the details of the characteristics and inclusion or exclusion
criteria for each work mentioned, providing a more precise and structured
description of the key aspects evaluated in each study. In addition, a brief
discussion of the ethical implications has been included in our conclusion to
avoid redundancies throughout the text and improve the document’s coher-
ence. These modifications aim to cover the topics of interest in bioelectrical
signals. Tables summarizing the content of the dataset used have been added,
along with brief descriptions of whether any type of inclusion exists within
each dataset.
Reviewer 2 Report
Comments and Suggestions for AuthorsIn your manuscript, you should include tables summarizing the approaches and results of the articles you have discussed, so that the reader can see the new trends you are using in your research.
It could possibly be pointed out that the very fast signal processing for rnn makes it difficult to generalize.
Also, the models that have been presented should mention how patient specific they are or are not. That is, how well you can create a generic model, or rather a personalized model.
Also the data set sizes should be indicated. For EEG data processing, there are already specific networks that have been created for this problem. Also hybrid networks should be mentioned.
In addition, an interesting aspect could be whether the cost function was modified or built-in during training.
Author Response
Reviewer 2
Comment 1: Include tables summarizing the approaches and results of the articles dis-
cussed, so the reader can see the new trends in research.
Response:
We greatly acknowledge the provided comment. In response, tables summariz-
ing the key aspects of each paper, the methodologies, and the metrics results
of the articles discussed throughout the manuscript have been added. The ta-
bles intend to provide a clearer overview of the papers and to facilitate the
reader’s understanding of the main contributions of recent research. Tables
can be found in each subsection of each signal.
Comment 2: It could possibly be pointed out that the very fast signal processing for RNN
makes it difficult to generalize.
Response:
Thank you for the valuable comment. It is agreed that despite the effectiveness
of RNNs for fast signal processing, they are often a limitation to be taken into
account, since this feature can pose generalization problems, hence a brief
explanation of this specific limitation has been included to solve it, as well as
an alternative to ensure the generalization efficiency of the model.
Comment 3: The models that have been presented should mention how patient-specific
they are or are not. That is, how well you can create a generic model, or
rather a personalized model.
Response:
Thank you for your thoughtful comment. The literature has been reviewed
and the models developed and datasets used have been reviewed, adding spe-
cific explanations to each paper to indicate whether the dataset used was for
specific patients or a generalizable dataset was used, that is, it was empha-
sized whether the data used are specific data or not, as well as discussing the
4potential implications of model customization with respect to the use of the
data used.
Comment 4: The dataset sizes should be indicated. For EEG data processing, there are
already specific networks that have been created for this problem. Also hybrid
networks should be mentioned.
Response:
Thank you for your valuable suggestions. The manuscript has been revised to
include the dataset sizes, as well as the types of data the datasets consisted
of for the reviewed studies, whenever the information was available. In ad-
dition, an analysis of networks designed specifically for signal processing was
included, specifically for EEG signal processing; these points are addressed in
each paper in each specific section.
Comment 5: An interesting aspect could be whether the cost function was modified or
built-in during training.
Response:
The manuscript has been revised to include the information in the studies
where applicable and the built-in cost functions were mentioned or custom
functions were implemented during training. Most of the research used stan-
dard cost functions such as cross entropy or mean square error, for those
papers that used specifically designed functions small descriptions were added
about what they consisted of.
Reviewer 3 Report
Comments and Suggestions for AuthorsDear authors,
The article is well designed and organized, however it presents some flaws that need underline. In particular, some either formal or content flaws should be ameliorated:
The title is obvious but it is absent in terms of the type of signals to compare. Also, the abstract lacks the objectives and methodology to carry out; and it lacks the subject before ‘Comparing’.
It shows duplicated referencing e.g. Eldele et al. [32]: please apply a single reference style;
An organised framework of comparison between distinct bioelectrical signals type is missing.
The article needs a ‘limitations’ subsection, maybe in the Conclusion to address AI limitations e.g. differences in terms of anatomy and ensuing signal strength.
Best regards,
Comments on the Quality of English LanguageMinor grammar errors
Author Response
Reviewer 3
Comment 1: The title is obvious but lacks the type of signals to compare. The abstract
lacks the objectives and methodology; it also lacks the subject before ‘Com-
paring’.
Response:
We appreciate your valuable comments on the title and summary of our re-
search. We would like to clarify that we decided not to specify the type of
signals in the title to avoid making it overly long. However, in the abstract,
we have included the signals analyzed throughout the research, as well as the
objective and methodology employed. Additionally, Figure 1, presented in the
introduction, describes the signals of interest addressed in the study in detail.
Comment 2: It shows duplicated referencing (e.g., Eldele et al. [32]): please apply a single
reference style.
5Response:
Thank you for pointing out the format of the citations. We decided to mention
the first author and use the numbering format for references in the text and
tables.
Comment 3: An organised framework of comparison between distinct bioelectrical signal
types is missing.
Response:
We greatly acknowledge the provided comment. In response, tables summariz-
ing the key aspects of each paper, the methodologies, and the metrics results
of the articles discussed throughout the manuscript have been added. The ta-
bles intend to provide a clearer overview of the papers and to facilitate the
reader’s understanding of the main contributions of recent research. Tables
can be found in each subsection of each signal.
Comment 4: The article needs a ‘limitations’ subsection, maybe in the Conclusion, to
address AI limitations e.g., differences in terms of anatomy and ensuing signal
strength.
Response:
Thank you for your valuable suggestion regarding including a ’limitations’
subsection. We have addressed this concern by incorporating the limitations
associated with using AI in each bioelectrical signal within the respective sec-
tions of the article. We have also provided a summary table in the Con-
clusion, highlighting the main differences between the signals and how these
differences impact research outcomes. This addition aims to provide a more
comprehensive understanding of the challenges and limitations of applying AI
to bioelectrical signal processing.
Reviewer 4 Report
Comments and Suggestions for Authors1. The Abstract should include some key findings, methodological recommendations, and precise mentions of the conclusions drawn from the review.
2. The statement 'When a cell (for our case study, a neuron) receives a stimulus, the membrane’s electrical potential changes rapidly, involving three main processes: Resting potential, Depolarization, and Repolarization.' is more of a physical description rather than a direct discussion of signal measurement. What is the specific reason for including this explanation?How do these characteristics impact signal measurement?
4.This paper summarizes different AI techniques applied to various physiological signals. However, factors such as experimental design, application domain, dataset size, preprocessing inclusion, and the use of transfer learning and fine-tuning can all impact accuracy. Additionally, whether the task is a classification problem (including the number of classes) or a regression problem also affects performance. These aspects have not been discussed in the paper.
Author Response
Reviewer 4
Comment 1: The Abstract should include key findings, methodological recommendations,
and precise mentions of the conclusions.
Response:
Thank you for your helpful suggestion. The Abstract has been revised to make
it more complete by including key results of the review, the methodological rec-
ommendations derived from each analysis and a clearer summary of the main
conclusions. By addressing this comment and implementing these changes, it
is intended to provide the reader with a more comprehensive and informative
overview of the article in a concise manner.
6Comment 2: The statement on neuron activity is more of a physical description. What
is the reason for including this? How do these characteristics impact signal
measurement?
Response:
The introduction of a bioelectrical signal by the physiological process was ini-
tially approached as a fundamental interpretation of how bioelectrical signals
are generated, as this process is inherent to all signals. However, instead of
addressing the entire physiological process in detail, it was chosen to sum-
marize it, focusing on the key aspects of bioelectrical signal generation. This
approach allows for a concise overview without needing to delve into the spe-
cific physiological elements of each signal, as the process is common to all
types of bioelectrical signals. Based on this fundamental explanation, the pri-
mary problems and sources of artifacts in bioelectric signals are discussed.
Comment 3: The paper summarizes AI techniques applied to physiological signals. How-
ever, factors like experimental design, domain, dataset size, preprocessing,
use of transfer learning and fine-tuning, and task type (classification vs. re-
gression) all affect performance and should be discussed.
Response:
The discussion in the manuscript was expanded to include the key factors in
the performance of the models for each signal work. The initial conditions of
each investigation, the description of the database used, the type of model de-
veloped, a discussion of the limitations of each model with respect to all of its
characteristics, and a discussion of the limitations of each model were added
and integrated into each section and job analyzed, as well as the key points
of each investigation in the tables to provide a more complete understanding
to the reader.
Round 2
Reviewer 1 Report
Comments and Suggestions for AuthorsThe authors have addressed my comments; therefore, the paper can be accepted for publication in the present format.
Reviewer 4 Report
Comments and Suggestions for AuthorsAuthors revised the comments according the reviwer's comments.