Review Reports - An Incremental Denoising Convolutional Autoencoder for Adaptive One-Dimensional Biosignal Reconstruction in Wearable Devices

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The manuscript addresses an important problem related to adaptive biosignal processing for resource-constrained wearable devices. The authors propose an incremental learning framework based on a denoising convolutional autoencoder (D-CAE) combined with Elastic Weight Consolidation (EWC), and evaluate it using ECG, EMG, and EEG datasets. The experimental results indicate potentially useful improvements in reconstruction metrics when comparing the incremental I-D-CAE model with the baseline D-CAE. However, in its current form, the manuscript does not yet provide sufficient experimental evidence for several of its central claims. Some methodological and terminology issues also need to be clarified.

Major Comments

The title and manuscript refer to the proposed approach as “Deep Incremental Metric Learning” but the presented optimization appears to consist primarily of D-CAE reconstruction using an MSE objective combined with an EWC regularization term during incremental retraining. It is hard to identify which is the conventional metric-learning objective. The authors should clearly explain what makes the proposed approach specifically “metric learning”. If “metric” refers to updating learned weights and biases, this terminology may be misleading.
EWC is introduced specifically to preserve previously learned knowledge, and the manuscript states that the proposed framework maintains performance on historical and newly acquired data. However, the results do not appear to provide the standard experiment required to demonstrate this property. Ideally, the authors should report a task-by-task performance matrix after each incremental learning session. Comparisons: independent/random initialization, sequential fine tuning without EWC, sequential fine tuning with EWC.
The manuscript emphasizes memory efficiency and suitability for wearable devices. However, maintaining the same model parameter size after incremental updating does not by itself demonstrate memory-efficient adaptation. The authors should clarify the memory requirements during incremental training, including any additional storage required for previous parameters, Fisher information, intermediate data, or optimizer states. Moreover, the experimental setup uses Windows 10 on an Intel i5 system with 8 GB RAM, which is substantially different from a resource-constrained wearable processor. Therefore, statements suggesting that the proposed approach is already suitable for wearable implementation should either be moderated or supported by additional resource analysis or experiments on representative edge/wearable hardware. At minimum, peak RAM consumption, storage requirements, and energy consumption or estimated energy cost should be discussed.
Most results are presented as single numerical values. The manuscript does not appear to provide standard deviations, confidence intervals, repeated experiments with different random seeds, or statistical significance tests. Deep-learning results can vary considerably depending on initialization and data partitioning. The authors should perform multiple independent runs and report, at minimum, mean and standard deviation for the main metrics. This is especially important for assessing whether the reported differences between D-CAE and I-D-CAE are robust rather than dependent on a particular training run.

Minor Comments

The manuscript would also benefit from a careful editorial revision. Several terminology and formatting inconsistencies are present. For example “IDCAE” and “I-D-CAE”, “ME-DIML” and “DIML”.
There also appears to be duplicated subsection numbering: Section 5.4.1 is first used for the general performance analysis and later again for the D-CAE versus I-D-CAE comparison.
The first page contains a template inconsistency: the manuscript is formatted as “Bioengineering”, while the footer still shows “Sensors”.

Author Response

Response:
We thank the reviewer for the valuable time and constructive feedback that has truly helped to improve the article. We have tried to incorporate all the suggested changes to the best of our ability in the revised article.

Major Comments

Comment 1:

The title and manuscript refer to the proposed approach as “Deep Incremental Metric Learning” but the presented optimization appears to consist primarily of D-CAE reconstruction using an MSE objective combined with an EWC regularization term during incremental retraining. It is hard to identify which is the conventional metric-learning objective. The authors should clearly explain what makes the proposed approach specifically “metric learning”. If “metric” refers to updating learned weights and biases, this terminology may be misleading.

Response:
We thank the reviewer for identifying this important terminology issue. We thank the reviewer for identifying this important terminology issue. We agree that the original term “Deep Incremental Metric Learning” was potentially misleading because the proposed method does not employ a conventional metric-learning objective, such as contrastive loss, triplet loss, or a learned embedding-space distance function.

The proposed approach is now consistently referred to as the Incremental Denoising Convolutional Autoencoder (I-D-CAE), which more accurately reflects the implemented methodology. The manuscript has also been revised to clarify that the proposed framework performs incremental reconstruction learning by updating previously learned model parameters using newly available data, with the reconstruction loss and EWC regularization used during incremental training.

To ensure consistency, the revised terminology has been applied throughout the manuscript, including the title, abstract, main text, equations, figures, tables, and captions. The revised title is:

“An Incremental Denoising Convolutional Autoencoder for Adaptive One-Dimensional Bio signal Reconstruction in Wearable Devices.”

The abstract has also been revised accordingly to describe the proposed method as an incremental learning framework based on I-D-CAE and to focus specifically on adaptive biosignal reconstruction rather than metric learning. These changes more precisely represent the implemented methodology and address the reviewer’s concern regarding the use of the term “metric learning.”.

Comment 2:

EWC is introduced specifically to preserve previously learned knowledge, and the manuscript states that the proposed framework maintains performance on historical and newly acquired data. However, the results do not appear to provide the standard experiment required to demonstrate this property. Ideally, the authors should report a task-by-task performance matrix after each incremental learning session. Comparisons: independent/random initialization, sequential fine tuning without EWC, sequential fine tuning with EWC.

Response:
We agree that demonstrating preservation of previously acquired knowledge is essential for validating the use of EWC. The revised manuscript therefore provides a clearer incremental-learning evaluation in which performance is examined on both previously observed and newly introduced data after each incremental learning stage. Table 13 and 14 are revised in the revised article and the text is also updated accordingly in the revised article.

Comment 3:

The manuscript emphasizes memory efficiency and suitability for wearable devices. However, maintaining the same model parameter size after incremental updating does not by itself demonstrate memory-efficient adaptation. The authors should clarify the memory requirements during incremental training, including any additional storage required for previous parameters, Fisher information, intermediate data, or optimizer states. Moreover, the experimental setup uses Windows 10 on an Intel i5 system with 8 GB RAM, which is substantially different from a resource-constrained wearable processor. Therefore, statements suggesting that the proposed approach is already suitable for wearable implementation should either be moderated or supported by additional resource analysis or experiments on representative edge/wearable hardware. At minimum, peak RAM consumption, storage requirements, and energy consumption or estimated energy cost should be discussed.

Response:
We agree that maintaining a fixed number of model parameters alone is not sufficient to establish memory-efficient incremental learning. Accordingly, we have revised the manuscript to clearly distinguish between model-size efficiency and the additional memory requirements associated with incremental training. In the proposed framework, the deep learning model is initially trained offline using a comparatively resource-rich computing platform, where the optimal weights and biases are learned. These learned parameters are subsequently used as the starting point for incremental updating when new data become available. This design is motivated by the fact that the statistical characteristics of physiological data may vary across patients and over time, potentially requiring periodic model adaptation. Rather than retraining the complete network from randomly initialized parameters, the proposed I-D-CAE framework updates the previously learned parameters using newly available data while preserving the network architecture and parameter count.

We also acknowledge that maintaining a fixed model parameter size does not necessarily imply an unchanged peak memory requirement during incremental training, as the training process may require additional memory for intermediate activations, gradients, newly acquired training data, and other training-related states. The incremental training process may additionally require memory for the previous model parameters, newly acquired training samples, intermediate activations, gradients, and, depending on the optimization strategy, optimizer states. In the case of EWC-based updating, the Fisher information and the corresponding reference parameters also introduce additional memory requirements. These training-time resources are distinct from the memory required to store the final inference model on the wearable device.

Therefore, the wearable-device suitability claimed in the article is intended primarily in terms of the deployment/inference model size and the avoidance of model-parameter expansion after incremental updating. The experiments were conducted on a Windows 10 system with an Intel i5 processor and 8 GB RAM to evaluate the proposed learning methodology and reconstruction performance. We acknowledge that this computing environment is not representative of a resource-constrained wearable processor. A hardware-level evaluation, including peak RAM consumption, storage overhead, execution latency, and energy consumption, would be required to establish the feasibility of fully on-device incremental training. Consequently, the revised article explicitly identifies the absence of direct measurements of peak RAM usage, energy consumption, and execution characteristics on representative wearable/edge hardware as a limitation. To provide clarity on the limitations in the study, we have moderated the corresponding statements in the revised manuscript to avoid implying direct on-device training without dedicated hardware validation. We have also revised the manuscript accordingly to differentiate fixed inference-model storage from incremental-training memory overhead and to ensure that our claims regarding wearable-device suitability are appropriately scoped.

Comment 4:

Most results are presented as single numerical values. The manuscript does not appear to provide standard deviations, confidence intervals, repeated experiments with different random seeds, or statistical significance tests. Deep-learning results can vary considerably depending on initialization and data partitioning. The authors should perform multiple independent runs and report, at minimum, mean and standard deviation for the main metrics. This is especially important for assessing whether the reported differences between D-CAE and I-D-CAE are robust rather than dependent on a particular training run.

Response:
We acknowledge that reporting a single numerical result does not quantify experimental variability. The proposed incremental protocol initializes each stage from the parameters learned in the preceding stage rather than from independent random initialization. Therefore, the reported results correspond to a fixed sequential training protocol. We have clarified this experimental design in the revised manuscript and have avoided interpreting the reported single-run values as estimates of population-level variability or statistical significance. We also acknowledge the absence of repeated independent runs and statistical significance testing as a limitation of the present study.

Minor Comments

Comment 1:

The manuscript would also benefit from a careful editorial revision. Several terminology and formatting inconsistencies are present. For example, “IDCAE” and “I-D-CAE”, “ME-DIML” and “DIML”.

Response:
The manuscript has been systematically revised to ensure consistent terminology throughout the text. The proposed model is now consistently referred to as I-D-CAE (Incremental Denoising Convolutional Autoencoder), while ME-DIML is consistently referred to as DIML. This revision eliminates terminological inconsistencies and ensures that the nomenclature remains clear and uniform throughout the manuscript.

Comment 2:

There also appears to be duplicated subsection numbering: Section 5.4.1 is first used for the general performance analysis and later again for the D-CAE versus I-D-CAE comparison.

Response:
Thank you for pointing out the error. The subsection numbering has been corrected, and the section hierarchy has been checked throughout the revised manuscript.

Comment 3

The first page contains a template inconsistency: the manuscript is formatted as “Bioengineering”, while the footer still shows “Sensors”.

Response:
The manuscript has been reformatted using the appropriate journal template, and the conflicting journal/template references have been removed.

Reviewer 2 Report

Comments and Suggestions for Authors

This manuscript presents a relevant study that adapts existing models when new data is introduced without retaining the entire historical dataset. It also adapts to different biosignals such as ECG, EMG and EEG. However, some considerations should be taken before it is accepted. These include:

Three different datasets are used to validate the data. However, the author did not specify the types of data. Are they numerical or continuous signals? If they differ, how can the authors ensure the data are suitable for the same metrics?
The author uses metric-learning objectives. However, for validation, the author uses losses and model initialization. This is not a metric-based learning. Therefore, the author should explain precisely what makes this a metric-learning method
Some minor issues like:

The abbreviations are inconsistent. ME-DIML, DIML, D-CAE, I-D-CAE, and IDCAE are used in different places. One naming convention should be adopted
Section 5.4.1 is duplicate
The author should use bullet points rather than "o" in the introduction
Some sentences are grammatically incorrect. Suggest using a full English proofread

Comments on the Quality of English Language

Author Response

Response:
We thank the reviewer for the valuable time and encouraging feedback. We have tried to incorporate all the suggested changes to the best of our ability in the revised article.

Major comments

Comment 1:

Three different datasets are used to validate the data. However, the author did not specify the types of data. Are they numerical or continuous signals? If they differ, how can the authors ensure the data are suitable for the same metrics?

Response:
We thank the reviewer for raising this important clarification. ECG, EMG, and EEG are all continuous physiological signals that are acquired and represented computationally as sampled numerical time-series data. Although these signals differ in their physiological origins, amplitude characteristics, frequency ranges, and temporal patterns, they can be represented in a common numerical form suitable for processing by the proposed deep learning framework. To ensure consistency across the three datasets, the input signals are preprocessed and normalized to the range of [−1,1]. This normalization places the signals on a common numerical scale while preserving their relative temporal characteristics. The same preprocessing and normalization procedure is applied before training and evaluation so that the input and reconstructed signals are directly comparable.

The use of ECG, EMG, and EEG is intended to evaluate the applicability and robustness of the proposed incremental reconstruction framework across different types of physiological biosignals, rather than to assume that their underlying physiological characteristics are identical. Although their statistical and spectral properties differ, the proposed model processes them in a common numerical representation. The reconstruction metrics are calculated between the corresponding normalized original and reconstructed signals. Therefore, the metrics evaluate the reconstruction error on the same signal domain and scale for each dataset. The normalization also prevents differences in the absolute amplitude ranges of ECG, EMG, and EEG from disproportionately influencing the reconstruction-error measurements. Thus, while the three datasets represent different physiological signal modalities, their common normalized numerical representation enables the proposed framework and reconstruction metrics to be applied consistently. Information about this is also added in the revised article.

Comment 2:

The author uses metric-learning objectives. However, for validation, the author uses losses and model initialization. This is not a metric-based learning. Therefore, the author should explain precisely what makes this a metric-learning method.

Response:
We thank the reviewer for identifying this important terminology issue. We agree that the original manuscript did not sufficiently distinguish conventional metric learning from the proposed incremental reconstruction-learning framework. In the revised manuscript, we clarify that the core optimization objective consists of the reconstruction loss of the Denoising Convolutional Autoencoder (D-CAE), together with the Elastic Weight Consolidation (EWC) regularization term. The term “metric learning” was therefore potentially misleading, as the proposed method does not employ a conventional distance-learning objective, such as contrastive loss, triplet loss, or a learned embedding-space distance function. To eliminate this ambiguity, we have revised the terminology throughout the manuscript and now consistently refer to the proposed model as the Incremental Denoising Convolutional Autoencoder (I-D-CAE). The methodological formulation and corresponding discussion have also been revised to explicitly distinguish the proposed incremental reconstruction-learning approach from conventional metric-learning methods. The proposed I-D-CAE performs incremental reconstruction learning through sequential fine-tuning with EWC, rather than conventional metric learning. The revised terminology more accurately represents the implemented methodology and avoids implying the use of a learned distance or similarity metric. Accordingly, the revised manuscript no longer characterizes the proposed method as a metric-learning approach. These changes have been incorporated throughout the revised manuscript and the corresponding modifications have been highlighted.

Minor Comments

Comment 1:

The abbreviations are inconsistent. ME-DIML, DIML, D-CAE, I-D-CAE, and IDCAE are used in different places. One naming convention should be adopted.

Response:
We thank the reviewer for pointing out the inconsistency in the abbreviations. We agree that using multiple abbreviations for the same or related models may cause ambiguity and reduce the clarity of the manuscript. Accordingly, the manuscript has been systematically revised to adopt a single and consistent naming convention throughout. The proposed model is now consistently referred to as I-D-CAE (Incremental Denoising Convolutional Autoencoder). The abbreviations ME-DIML and IDCAE have been removed to avoid duplication and inconsistency. The revised terminology has been applied consistently throughout the main text, figures, tables, equations, and captions. These changes have been highlighted in the revised manuscript.

Comment 2:

Section 5.4.1 is duplicate

Response:
Thank you for highlighting the error. The duplicate subsection numbering has been removed and the section numbering has been checked throughout the manuscript.

Comment 3:

The author should use bullet points rather than "o" in the introduction.

Response:
We thank the reviewer. The formatting of lists has been corrected using consistent bullet points. In addition, the manuscript has undergone a comprehensive English-language revision to correct grammatical, syntactic, terminology, and formatting issues.

Comment 4:

Some sentences are grammatically incorrect. Suggest using a full English proofread

Response:
Thank you for the suggestion. We have carefully proofread the manuscript and corrected the grammatical and language-related issues throughout the document.

Reviewer 3 Report

Comments and Suggestions for Authors

The search for structure in training multi-level neural network has produced deep learning: a concept to experiment with a limited number of circuit parts with better results. Unfortunately, it has not reached a search structure. Rather, a number of authors have merely found interesting solutions to claim their fame. Their fundamental problem is that no neural learning/network combination has ever been the very best under all circumstances, a “gratia probatum”. This is also true when all ingredients are named in a single designation, such as “DasaNet”.

Deep Learning is not limited to one-dimensional signals. It can be used for signal plots and for (a)synchronous time-sequenced ones, adding different dimensions to the already available variety in the structural domain. Basically, the manuscript simplifies the issue by merely comparing the author’s selection with the closest competition. A further limitation is by experimenting on a well-known medical signal as widely available in for instance the MIT bio-motor sets. The recognition of such a signal is usually aimed at qualification, not even on quantification, let alone on being precise.

The presented text is designed well. It has an introduction to the topic, a section of the mathematical background for suitable modeling and 1 or more experiments followed by conclusions. So far, I concluded that the title/introduction combination has eluded me into having no clear subject. I was wrong!! Given a set of signal-plots, perfect replication can be achieved. With a high enough sampling rate, the signal can always be restored. However, the bio-signal is constantly changing based on a stochastic motor, but typical abnormal points in the curve are clearly observable.

This ideal is far from practice. Real signals are very noisy, depending in non-controllable handling situations, and should not be filtered (as happened in the past). Coming to think of it, a signal plot can be closely approximated as reported in the proposed paper, or only correlated as in reality. Recognition of QRS-complexes are ranging from 5 to 25% error. Such numbers may not see impressive but they are at least realistic and can still be handled by deep learning.

The submitted manuscript is well designed, but the message can be more consistent and recognizable. The amount of literature references can be reduced once the message of the contribution has been clarified. A number of 25 references seems more than enough. The figures are interesting, but the value of the paper will not diminish with simplifying some figures. The model mathematics is spread out over several sections. Overall the composition of the paper is vaguely familiar to me. Again, I conclude “Less is More”. It pays to check what is actually needed to support the aims and scope. In all, well written but it has no body yet.

Author Response

The submitted manuscript is well designed, but the message can be more consistent and recognizable. The amount of literature references can be reduced once the message of the contribution has been clarified. A number of 25 references seems more than enough. The figures are interesting, but the value of the paper will not diminish with simplifying some figures. The model mathematics is spread out over several sections. Overall, the composition of the paper is vaguely familiar to me. Again, I conclude “Less is More”. It pays to check what is actually needed to support the aims and scope. In all, well written but it has no body yet.

We thank the reviewer for the valuable time and valuable feedback. We have tried to incorporate all the suggested changes to the best of our ability in the revised article.

Comment: Clarification of the central contribution and scope

Reviewer comment:

Response:
We sincerely thank the reviewer for the comprehensive and constructive comments concerning the scope of deep learning, the stochastic nature of physiological signals, the distinction between signal reconstruction and recognition, and the practical limitations of biosignal processing. We have carefully considered these observations and have revised the manuscript to more precisely define the technical contribution, experimental scope, and limitations of the proposed framework.

Scope and Objective of the Proposed Deep Learning Framework

We agree that deep learning is a general computational paradigm and is not restricted to one-dimensional signals. It can be applied to one-dimensional temporal sequences, two-dimensional signal representations, multidimensional physiological data, asynchronous sequences, and other structured domains. The objective of the present work is not to establish a universally optimal neural-network architecture or to claim that a particular architecture is superior under all possible operating conditions. The specific objective of this study is to develop an incremental deep-learning framework for biosignal reconstruction. The proposed Incremental Denoising Convolutional Autoencoder (I-D-CAE) learns a nonlinear mapping between the input and reconstructed biosignal representations and subsequently adapts the learned parameters when newly available signal data exhibit different statistical characteristics. Thus, the contribution is centered on incremental reconstruction learning and parameter adaptation, rather than on proposing a general-purpose deep-learning architecture. The manuscript has been revised to make this scope explicit and to avoid generalized claims regarding the superiority of the proposed architecture over all alternative deep-learning approaches.

Non-Stationarity and Stochastic Characteristics of Biosignals

We agree with the reviewer that physiological biosignals cannot generally be regarded as deterministic or stationary waveforms. Their statistical properties may vary as a function of subject, physiological state, recording session, sensor placement, motion, electrode impedance, instrumentation, and environmental interference. Consequently, a reconstruction model trained under one statistical distribution may experience a distribution shift when exposed to subsequently acquired data. This issue provides the motivation for the incremental-learning component of the proposed framework. Instead of independently retraining the complete network from randomly initialized parameters whenever new data become available, the I-D-CAE retains the previously learned parameter state and sequentially adapts the model using newly available data. EWC is incorporated to constrain changes to parameters that are important for previously learned information, thereby reducing the risk of catastrophic forgetting. Therefore, the proposed approach does not assume that a single offline-trained model is universally optimal. Rather, it is designed to provide a mechanism for progressive adaptation to evolving biosignal distributions while maintaining the existing model architecture and parameter dimensionality.

Distinction Between Signal Reconstruction and Signal Recognition

We particularly appreciate the reviewer’s distinction between signal reconstruction and signal recognition. These are fundamentally different problem formulations and should not be evaluated using the same objectives. Signal-recognition tasks, such as QRS-complex detection, arrhythmia classification, event detection, or disease classification, primarily evaluate whether clinically or physiologically meaningful patterns can be correctly identified. In contrast, the objective of the present study is to reconstruct the waveform itself as accurately as possible. Accordingly, the evaluation in this work focuses on quantitative reconstruction measures, including RMSE, PRD, SNR, QS, and compression-related measures, rather than classification accuracy or event-detection error. The reported improvements therefore indicate enhanced waveform reconstruction under the experimental conditions considered; they should not be interpreted as evidence of improved diagnostic or clinical classification performance. We have revised the manuscript to explicitly distinguish these objectives and to avoid conflating reconstruction fidelity with clinical recognition performance.

Reconstruction Fidelity Versus Perfect Physiological Signal Recovery

We also agree that a sufficiently high sampling rate can allow a waveform to be approximated with high numerical fidelity, but this does not imply that the reconstructed waveform represents a perfect recovery of the underlying physiological process. Physiological signals contain stochastic variations and may be affected by transient artifacts, sensor noise, motion artifacts, electrode displacement, baseline variations, and other uncontrolled factors. Therefore, the objective of the proposed I-D-CAE is not to claim perfect reconstruction of the underlying physiological state. Rather, it is to minimize the reconstruction error between the defined input/target signal representations within the experimental framework. The reconstructed output should consequently be interpreted as a learned approximation of the target biosignal rather than as an exact recovery of an unobservable physiological ground truth. This distinction has been explicitly incorporated into the revised manuscript.

Denoising and the Risk of Removing Physiologically Relevant Information

We acknowledge the important concern regarding conventional filtering and the possibility of suppressing physiologically meaningful signal components. We agree that aggressive filtering can alter waveform morphology or attenuate signal components that may subsequently be relevant for physiological interpretation. The proposed approach should therefore not be interpreted simply as a conventional deterministic filtering operation. The D-CAE learns a nonlinear reconstruction mapping from the training data, with the reconstruction objective determining the relationship between the input and target representations. The purpose is to learn the statistical structure required for reconstruction rather than to impose a manually selected frequency-domain filtering rule. We also acknowledge that any learned denoising/reconstruction model can potentially suppress information that is poorly represented in its training distribution. Consequently, the revised manuscript explicitly identifies the dependence of reconstruction performance on the characteristics and representativeness of the training data as an important limitation.

Multi-Modal Biosignal Evaluation

The use of ECG, EMG, and EEG in this study is intended to examine the applicability of the proposed incremental reconstruction mechanism across different physiological signal modalities. These signals have substantially different physiological origins, frequency characteristics, amplitude distributions, temporal structures, and noise characteristics. We do not assume that these modalities are physiologically equivalent. Instead, they provide distinct test cases for evaluating whether the proposed learning and incremental adaptation mechanism remains effective when applied to different classes of biosignals. As reported in Table 13, the I-D-CAE provides improved reconstruction performance relative to independently trained D-CAE models across all three signal modalities. For example, RMSE decreases from 0.021 to 0.0033 for ECG, from 0.0055 to 0.0011 for EMG, and from 0.0025 to 0.0008 for EEG. The corresponding SNR values increase from 25.70 to 29.71 dB for ECG, from 24.45 to 35 dB for EMG, and from 18.76 to 30.8 dB for EEG.

These results demonstrate the effectiveness of the proposed incremental reconstruction mechanism under the evaluated experimental conditions. However, we have deliberately avoided interpreting these results as evidence of universal superiority across all possible biosignal datasets.

Incremental Adaptation and Computational Considerations

The principal motivation for incremental learning is that biosignal distributions encountered during subsequent acquisition may differ from those represented during the initial offline training stage. Conventional independent training requires the model to be retrained as a separate learning process, whereas the proposed sequential strategy uses the previously learned parameter state as the starting point for subsequent adaptation. Table 14 demonstrates that sequential fine-tuning with EWC reduces the measured training time relative to independent training. The training time decreases from 4305 to 3917 s for ECG, from 4725 to 4099 s for EMG, and from 3914 to 3581 s for EEG. More importantly, the learned weight-and-bias storage requirement remains unchanged after incremental updating: 954 kB for ECG, 947 kB for EMG, and 1125 kB for EEG. Thus, the proposed approach provides parameter-storage invariance at the model level, although we acknowledge that this should not be interpreted as proof that the complete incremental-training process has identical peak memory or energy requirements.

Limitations Regarding Real-World and Clinical Deployment

We acknowledge that experiments using publicly available biosignal datasets cannot fully reproduce the variability encountered in real-world wearable sensing environments. Factors such as subject-to-subject variability, electrode displacement, motion artifacts, sensor characteristics, environmental interference, physiological state, and long-term temporal drift may produce distributions that differ from those represented in the experimental datasets.

Furthermore, the present study does not perform downstream clinical validation of the reconstructed signals. In particular, we do not claim that improvements in RMSE, PRD, SNR, or QS directly translate into improved diagnostic accuracy, QRS detection, arrhythmia classification, or other clinical outcomes. Establishing such a relationship would require additional experiments in which clinically relevant downstream tasks are performed using both original and reconstructed signals. We have therefore revised the manuscript to explicitly identify real-world signal variability, broader cross-subject validation, downstream clinical validation, and hardware-level evaluation as areas requiring further investigation.

Clarification of the Contribution

Based on the reviewer’s comments, we have refined the presentation of the contribution. The proposed work should be understood as an incremental biosignal reconstruction framework based on an Incremental Denoising Convolutional Autoencoder (I-D-CAE) with sequential fine-tuning and EWC.

The principal contribution is not the claim of a universally optimal deep-learning architecture, nor the claim of perfect physiological signal recovery. Rather, the contribution is the development and evaluation of a mechanism that allows a previously trained reconstruction model to adapt sequentially to newly available biosignal data while retaining previously learned parameters and maintaining the model's parameter-storage size.

These clarifications and limitations have been incorporated into the revised manuscript to provide a more technically precise description of the proposed methodology, its experimental evidence, and its practical scope.

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

The authors have improved the manuscript by revising the title and abstract, clarifying the model comparison, and acknowledging the limitations of hardware-level validation. I recommend minor revision, provided that the remaining claims are aligned with the available experimental evidence.

Tables 13 and 14 support a comparison between independent training and sequential fine-tuning with EWC, but they do not directly demonstrate preservation of performance on previously learned tasks. Please moderate the corresponding statements.
Please check the consistency of the numerical results. In eq.(9) you obtained 38.56 dB for PRD = 1.18%, whereas Table 13 reports 35 dB for EMG. For EEG, PRD = 3.23% gives approximately 29.82 dB, compared with the reported 30.8 dB. Please correct these values or explain any differences in aggregation.

Finally, please remove the remaining metric-learning terminology from Figure 1 caption and any other sections if this is the case.

Author Response

Tables 13 and 14 support a comparison between independent training and sequential fine-tuning with EWC, but they do not directly demonstrate preservation of performance on previously learned tasks. Please moderate the corresponding statements.

Response: We thank the reviewer for this important observation. We agree that the results presented in Tables 13 and 14 do not directly establish the preservation of performance on previously learned tasks. But Section 5.3 clearly explains how the data is trained incrementally (sequential fine tuning) with EWC. Accordingly, we have revised the Table 13 and 14 to show the comparison between data trained with randomly initialized weight and sequential fine tuning of the data. The discussion now focuses on the observed comparison between independent training and sequential fine-tuning with EWC, while acknowledging that a direct assessment of performance retention on previously learned tasks would require additional evaluation. We believe this revision provides a more accurate interpretation of the experimental findings.

Please check the consistency of the numerical results. In eq.(9) you obtained 38.56 dB for PRD = 1.18%, whereas Table 13 reports 35 dB for EMG. For EEG, PRD = 3.23% gives approximately 29.82 dB, compared with the reported 30.8 dB. Please correct these values or explain any differences in aggregation.

Response: We thank the reviewer for identifying these numerical inconsistencies. We have re-examined the relationship between percentage root-mean-square difference (PRD) and the corresponding signal-to-noise ratio (SNR), as well as the values reported in Eq. (9) and Table 13. We agree that the reported values require consistency checks. The revised manuscript corrects the corresponding numerical values where necessary and clarifies the calculation and aggregation procedures to ensure that the reported PRD and SNR values are interpreted consistently. Any differences arising from the use of individual-record results versus aggregated results are explicitly clarified, where applicable.

Finally, please remove the remaining metric-learning terminology from Figure 1 caption and any other sections if this is the case.

Response: We have revised the caption of Figure 1 and the relevant sections of the manuscript and removed the remaining metric-learning terminology where it was inconsistent with the methodology presented. The terminology has been revised to ensure consistency with the actual learning framework and experimental procedures described in the manuscript.

We thank the reviewer again for these valuable comments, which have helped us improve the accuracy, clarity, and consistency of the revised manuscript.

Reviewer 2 Report

Comments and Suggestions for Authors

The revised manuscript has improved the discussion of incremental learning and the issues on the metric learning method. However, I would suggest that the authors added more details on the hyper-parameter selection. In addition, it is unclear on the splitting process whether it is performed before or after segmentation. Or whether the split was random.

Comments on the Quality of English Language

Author Response

Response: We sincerely thank the reviewer for the positive assessment of the revised manuscript and for the valuable suggestions regarding hyperparameter selection and the data-splitting procedure. We agree that providing these details is important for improving the reproducibility and methodological clarity of the study. We have addressed the comments as follows.

(a) Hyperparameter selection.

We have expanded the description of the hyperparameter selection procedure in the revised manuscript. Section 4.2 clarifies the selection of the key hyper-parameters, including the number of training epochs. We have also clarified the criteria used to determine the final parameter settings. These revisions provide a clearer explanation of the model configuration and improve the reproducibility of the proposed approach.

(b) Data splitting

We have clarified the order of data preprocessing, segmentation, and dataset partitioning in the revised manuscript. We have also specified the splitting strategy and whether randomization was applied, together with the relevant proportions of the training and testing sets, as applicable in the Section 5.3.

We thank the reviewer again for these constructive comments. The revisions improve the transparency and reproducibility of the experimental methodology and make the implementation details clearer to readers.

Reviewer 3 Report

Comments and Suggestions for Authors

The question is still: what is it all about? Many words that are used in modern technology are also placed in the title. This ensures that any potentially question can be answered by citing some of these words. The assumption is that you can always be right once selecting the proper keywords. Vice versa, any paper will be right when enough keywords are sampled. Or even worse, there is no need for new reviews when the variance of the sampling is big enough. The time is come of the Review Generator to judge a given text.

This is essentially a matter for the Editor-in-Chief to decide. The criteria are good structure fitting to the education system, it is on the technical soundness measured in educational value. Furthermore, it is to be exemplary beyond being original, and so on. A number of these aspects are mentioned in reviewing forms, but this can never be complete. But even worse, it is questionable whether the required quantification can be achieved given the applied number of keywords. This almost sounds as being part of the explosively growing interest for misused Super Intelligence.

The provided text started me to question the current definition of research fraud. It is not sufficient to find overlap with previously published papers. On the one hand, it is required to quantify duplication, preferable in combination with lack smoothly growing understanding by the authors. On the other hand, jumps in understanding needs to allow evolution in combination with revolution. There is a lot of theoretical discussion on this topic. What we need foremost here is nuggets of personal growth that are unique and personally claimed.

I have not found such nuggets in the proposed paper. Consequently, I found nothing to review. Technically, there is nothing wrong but just missing a proper embodiment in the world-wide richness of published papers. There is nothing claimed to clearly indicated improvement beyond the state-of-the-art. In other words: there is no meat. And where there is no meat, there is also nothing to chew on. I have really tried to find the nuggets, but I found no gold to shine with. I suggest that digging deeper, but in a smaller area, is a good step forward. The current manuscript has simply tried to mask the emptiness by overgrowth. Even a Super Intelligent text compiler can do this, and in less time.

All-in-All, I have refrained from going into details of the proposed circuit scheme as this would require a small tutorial. Looking at the reaction of the authors leading to the current version, the main critiques on the original manuscript are largely ignored. Hence a real breakthrough in the temporariness of the presentation cannot be expected. Therefore, I have just outlined the principles of my critique and have not touched the remaining content. There is clearly still a lot of work to be done before a reviewer can really do his job.

Author Response

Response: We respectfully submit that the primary contribution of this work may not have been sufficiently articulated in the previous version. The objective of this study is not to claim a fundamentally new convolutional autoencoder architecture or to present sequential fine-tuning and elastic weight consolidation (EWC) as entirely new techniques. Rather, this work investigates their integration into an incremental denoising convolutional autoencoder framework for adaptive biosignal reconstruction, with particular emphasis on maintaining a fixed network architecture and model-parameter storage size while adapting to newly available data.

The motivation for this research arises from the practical challenge of adapting biosignal reconstruction models to changing physiological data distributions in wearable applications. Conventional retraining may require repeated access to historical datasets, whereas sequential fine-tuning enables model updates using previously learned parameters as the initialization for subsequent training. However, parameter initialization alone does not guarantee the preservation of previously acquired knowledge. To address this limitation, the proposed framework incorporates EWC to constrain changes to parameters that are important for previously learned data, thereby seeking to mitigate catastrophic forgetting during incremental adaptation.

The proposed approach has been evaluated using ECG, EMG, and EEG datasets. In particular, Table 13 compares independent training with sequential fine-tuning incorporating EWC, while Table 14 examines training time and model-parameter storage. The reported results indicate improved reconstruction metrics for the evaluated incremental-learning experiments while maintaining the same reported model-parameter storage size. We believe that these findings provide a practical experimental investigation of incremental biosignal reconstruction under a fixed model-parameter budget.

We would be pleased to revise the manuscript wherever necessary to address the reviewer’s concerns, strengthen the technical discussion, and clarify the scientific contributions of the work. We remain committed to incorporating appropriate revisions and additional clarifications, as required, to improve the manuscript’s rigor, clarity, and overall scientific quality.

We thank the reviewer again for encouraging us to examine the manuscript’s contribution and supporting evidence more critically. We believe that the revisions provide a more focused and transparent account of the study, its experimental findings, and its current limitations.