Investigation of the MobileNetV2 Optimal Feature Extraction Layer for EEG-Based Dementia Severity Classification: A Comparative Study

Comments 1: The relevance of using MobileNetV2, a lightweight 53-layer deep convolutional neural network model, in EEG analysis is beyond doubt. However, the concept of depth-separable convolutions, which apply one filter to each input channel and combine the output data using pointwise convolutions, when applied to EEG analysis requires a neurophysiological justification for the filters used. This paper does not provide a neurophysiological justification for the traditional filters used. In particular, the frequency filter from 1 to 70 Hz includes frequency ranges that have completely different functional significance for the diagnosis of EEG waves. In addition, the wide frequency range from 1 to 70 Hz includes EMG and ECG artifacts that pollute the EEG signal and distort the correctness of the diagnosis. In this regard, authors should explain how they solve these problems.

Response 1: We thank the reviewer for their insightful and valuable feedback regarding the application of MobileNetV2 in EEG analysis and the concerns raised about the choice of frequency filters and the handling of artifacts.

EEG signals are multichannel time-series data with spatial and temporal dependencies; depth-separable convolutions can be interpreted as a computationally efficient approach to capture channel-specific temporal patterns (depthwise step) and integrate cross-channel information (pointwise step). Neurophysiologically, EEG signals reflect the summed electrical activity of neuronal populations, with distinct frequency bands (e.g., delta, theta, alpha, beta, and gamma) corresponding to different cognitive and pathological states. The MobileNetV2 depthwise convolution can be considered analogous to extracting localized temporal features from individual EEG channels, which may correspond to specific oscillatory patterns or event-related potentials.

The reviewer correctly notes that the broad frequency range of 1–70 Hz encompasses multiple EEG frequency bands (delta: 1–4 Hz, theta: 4–8 Hz, alpha: 8–13 Hz, beta: 13–30 Hz, gamma: 30–70 Hz), each with distinct neurophysiological roles in cognitive and clinical contexts. For example, delta waves are associated with pathological states, while gamma waves are linked to cognitive processing. The choice of a 1–70 Hz bandpass filter was motivated by the need to capture a comprehensive range of EEG activity relevant to the diagnostic task (particularly dementia cognitive state classification). However, we acknowledge that this broad range may not be optimal for all applications and could include irrelevant or confounding signals. In our study, we applied standard preprocessing techniques, including notch filtering at 50 Hz to remove powerline noise; however, we recognize that EMG and ECG artifacts require more robust mitigation strategies. To address this, we have implemented additional preprocessing steps, including discrete wavelet transform (DWT) to remove noise associated with these artifacts. Additionally, we have included a discussion in Section 2.2 Preprocessing to address the wide tactile range of the 1–70 Hz range and to enhance signal quality. In the new version of the paper, major changes are highlighted in yellow.

“Additionally, the discrete wavelet transform (DWT) was one of several additional preprocessing procedures used to enhance data quality and remove artifacts throughout the wide tactile range of 0.1–70 Hz.”

Comments 2: Since the readers of MDPI journals are not only machine learning specialists, but also neurophysiologists-diagnostics, for whose sake the predictive models are created, authors should justify each step of their model in terms of neurophysiological adequacy.

each abbreviation at the beginning of the article should be deciphered. For example, the abbreviation for Discrete Wavelet transform (DWT) appears only in the conclusion.

Figure captions must be deciphered.

Response 2: Thank you once again for your valuable suggestions. We have made the necessary revisions to enhance the clarity and accessibility of our manuscript, ensuring it meets the needs of our diverse readership. In our revised manuscript, we have explicitly outlined these neurophysiological justifications for each methodological step, ensuring clarity for a diverse readership. We believe this strengthen the scientific rigor of our study and enhance its utility for both machine learning experts and neurophysiologists.

3. Additional clarifications

We appreciate your guidance and look forward to your feedback on our revised submission.

Comments 1: The final sentence of the abstract should briefly highlight the significance and contributions of the study.

Response 1: Thank you for your insightful comment regarding the final sentence of the abstract. We agree that emphasizing the significance and contributions of our study is crucial for conveying its importance. The revised version of the final sentence of the abstract as follows:

“The findings of this study endorse the utilization of deep processing algorithms, offering a viable approach for improving early dementia identification with high precision, hence facilitating the differentiation among NC individuals, VasD patients, and MCogImp patients.”

Comments 2: The paper should describe the rationale for the 70:30 split between the training and testing datasets. It is recommended to include results from alternative data split ratios for comparison.

Response 2: We thank the reviewer for the valuable suggestion to describe the rationale for the 70:30 split between training and testing datasets and to include results from alternative data split ratios for comparison. The 70:30 split, where 70% of the data is used for training and validation and 30% for testing, was selected following a preliminary evaluation of accuracy in our EEG-based study on dementia severity classification. We tested alternative ratios, with the 60:40 split yielding an accuracy of 66.67%, the 80:20 split achieving 73.7%, and the 70:30 split resulting in the highest accuracy of 74.9% using the MobileNetV2. Based on these results, we chose the 70:30 split, as it provided the best performance, reflecting an optimal balance between the sufficiency of training data and the reliability of the test set. Thus, we justify the 70:30 split based on both empirical evidence and literature support, as this ratio is also recommended for EEG classification tasks, where it has been shown to enhance model generalization (Rabcan et al. and Ganaie et al. reported good accuracies with a 70:30 split). To focus our resources on the core methodology, we did not pursue further exploration of other split ratios beyond this initial assessment. Instead, we prioritized a comparative study to investigate the optimal feature extraction layer of MobileNetV2 (convolutional layers ('Conv-1'), batch normalization ('Conv-1-bn'), clipped ReLU ('out-relu'), 2-D Global Average Pooling ('global-average-pooling2d1'), and fully connected ('Logits') layers) for EEG-based dementia severity classification, as detailed in Section 2.3 Deep Feature Extraction with MobileNetV2 Layers. This decision allowed us to deeply analyze layer-specific performance, which is the primary contribution of this work, rather than extensively varying split ratios. We have added the following paragraph to Section 2.4 Classification, and the major changes are highlighted in yellow in the revised manuscript.

“30% of the data is used for testing, and the remaining 70% is used for training and checking.  The 70:30 ratio is a good balance between training and testing data. It lets the model learn well while still being tested on enough amount of data. This ratio is commonly recommended in the literature [29,30] since it maintains generalization”

References:

29. Rabcan, Jan, et al. "EEG signal classification based on fuzzy classifiers." IEEE Transactions on Industrial Informatics 18.2 (2021): 757-766.

30. Ganaie, M. A., et al. "EEG signal classification using improved intuitionistic fuzzy twin support vector machines." Neural Computing and Applications 36.1 (2024): 163-179.

Comments 3: The classification performance metrics should be supported by appropriate references from relevant literature.

Response 3: We appreciate the reviewer’s insightful comment regarding the need to support the classification performance metrics with appropriate references from relevant literature. We have updated the manuscript to include citations that contextualize and validate the reported accuracies, with revisions incorporated into Section 2.4. Classification. In the new version of the paper, major changes are highlighted in yellow.

“The evaluation's scope is enlarged to include the accuracy and confusion matrix to evaluate dementia severity prediction performance [31]. These performance measures are chosen and interpreted based on generally accepted machine learning and deep learning principles [32].

As in medical diagnostics, the confusion matrix is an important tool for analyzing classification results since it provides a granular breakdown of model performance that goes beyond total accuracy. In the context of this EEG-based dementia classification study—differentiating NC, MCogImp, and VasD using MobileNetV2 features—its significance resides in analyzing class-specific performance and is compatible with best practices advised in the machine learning and deep learning studies [33,34].”

References:

31. Cherian, Resmi, and E. Gracemary Kanaga. "Theoretical and methodological analysis of EEG based seizure detection and prediction: An exhaustive review." Journal of neuroscience methods 369 (2022): 109483.

32. Lin, Feng, et al. "Predicting cognitive impairment in outpatients with epilepsy using machine learning techniques." Scientific Reports 11.1 (2021): 20002.

33. Hashmi, Arshad, and Omar Barukab. "Dementia classification using deep reinforcement learning for early diagnosis." Applied Sciences 13.3 (2023): 1464.

34. Ayman, Ummara, et al. "Epileptic patient activity recognition system using extreme learning machine method." Biomedicines 11.3 (2023): 816.

Comments 4: The process for selecting hyperparameters should be clearly explained, and the final values used should be reported in the manuscript.

Response 4: We thank the reviewer for the valuable suggestion to clearly explain the process for selecting hyperparameters and report the final values used in the manuscript. We have revised the manuscript to address this concern, with updates incorporated into Section 2.4. Classification. In the new version of the paper, major changes are highlighted in yellow.

“Additionally, the EEG classification model that relies on MobileNetV2 had its hyperparameters chosen using 70\% of the training dataset, which included the NC, MCogImp, and VasD groups.  Optimizations were made to the learning rate, batch size, and epoch count, which are critical hyperparameters.  A validation frequency of 3, a mini-batch size of 64, and a learning rate of $0.001$ were the final hyperparameters chosen.”

Comments 5: The study should incorporate comparisons with mainstream models such as XGBoost, SVR, and others to validate the performance of the proposed method.

Response 5: We appreciate the reviewer’s insightful suggestion to incorporate comparisons with mainstream models such as XGBoost, Support Vector Regression (SVR), and others to validate the performance of our proposed MobileNetV2-based method for EEG-based dementia classification.

We acknowledge the value of Our approach in benchmarking our achieved accuracy of 95.9% (using the Decision Tree classifier with global-average-pooling2d-1 features) against established algorithms. At present, we are focusing on a comparative study to identify the optimal feature extraction layer of MobileNetV2, which includes convolutional layers ('Conv-1'), batch normalization ('Conv-1-bn'), clipped ReLU ('out-relu'), 2-D Global Average Pooling ('global-average-pooling2d1'), and fully connected ('Logits') layers, for EEG-based dementia severity classification; further details can be found in Section 2.3, Deep Feature Extraction with MobileNetV2 Layers. However, we agree that comparing with mainstream models would strengthen the validation of our method. Given the current scope of the study, we have outlined this comparison as the next step in our research. In the revised manuscript, we have added a Future Work to Section 4. Conclusions detailing plans to implement XGBoost, SVR, and potentially other models like Random Forest on the same EEG spectrogram dataset, using the same preprocessing pipeline. We anticipate completing these comparisons in subsequent work and reporting the results in a follow-up publication. We have revised the manuscript to address this concern by adding the following paragraph to Section 4. Conclusions. In the new version of the paper, major changes are highlighted in yellow.

"Future work intends to use XGBoost, SVR, and possibly more models such as Random Forest on the identical EEG spectrogram dataset, utilizing the same preprocessing workflow."

Comments 6: The conclusion should briefly acknowledge the limitations of the study and outline directions for future research.

Response 6: We thank the reviewer for the constructive suggestion to briefly acknowledge the limitations of the study and outline directions for future research in the conclusion. We have revised the manuscript to address this concern, with updates incorporated into Section 4. Conclusions. In the new version of the paper, major changes are highlighted in yellow.

“The limitations of this study include the relatively small sample size (35 participants: 15 NC, 15 MCogImp, 5 VasD). The dependence on a singular working memory task may inadequately represent the varied EEG patterns linked to dementia progression, perhaps constraining the model’s relevance across numerous cognitive circumstances. Future work intends to use XGBoost, SVR, and possibly more models such as Random Forest on the identical EEG spectrogram dataset, utilizing the same preprocessing workflow."

5. Additional clarifications

We appreciate your guidance and look forward to your feedback on our revised submission.

Response to Reviewer 1 Comments

1. Summary

Thank you very much for taking the time to review this manuscript. We have carefully revised the manuscript following the Reviewers’ comments. We considered and addressed each one of their concerns and remarks.

Major changes are highlighted in yellow in the revised manuscript. Additionally, pieces of text that have been included in the revised manuscript to address the Reviewers’ comments appear in this response document typed in Italic font.

We really appreciate the Reviewer’s effort in revising our study. We have considered your comments thoroughly regarding the writing aspect. All of the revisions to the manuscript have been carefully considered.

We are grateful for the feedback provided by the Editor in Chief, Associate Editor, and Reviewers. Their remarks and suggestions helped us to improve the manuscript significantly. We hope that the revised version of the study has addressed all your concerns and will be considered a contribution of interest to the readership of “Algorithms Journal.”

For your convenience, a list of responses to the Reviewers’ remarks is included below.

2. Questions for General Evaluation

Reviewer’s Evaluation

Response and Revisions

Does the introduction provide sufficient background and include all relevant references?

Can be improved

Has been improved

Are all the cited references relevant to the research?

Yes

Thank you

Is the research design appropriate?

Yes

Thank you

Are the methods adequately described?

Must be improved

Has been improved

Are the results clearly presented?

Can be improved

Has been improved

Are the conclusions supported by the results?

Can be improved

Has been improved

3. Point-by-point response to Comments and Suggestions for Authors

Comments 1: No references to important studies on EEG-EMG signal coherence and its age dependence are provided.
James, L. M., Halliday, D. M., Stephens, J. A., & Farmer, S. F. (2008). On the development of human corticospinal oscillations: age‐related changes in EEG–EMG coherence and cumulant. European Journal of Neuroscience, 27(12), 3369–3379.
Mima, T. and Hallett, M., 1999. Corticomuscular coherence: a review. Journal of clinical neurophysiology, 16(6), p.501.
Grosse, P., M. J. Cassidy, and P. Brown. "EEG–EMG, MEG–EMG and EMG–EMG frequency analysis: physiological principles and clinical applications." Clinical Neurophysiology 113, no. 10 (2002): 1523-1531.
Liu, Jinbiao, Yixuan Sheng, and Honghai Liu. "Corticomuscular coherence and its applications: a review." Frontiers in human neuroscience 13 (2019): 100.
Tuncel, D., Dizibuyuk, A. and Kiymik, M.K., 2010. Time frequency based coherence analysis between EEG and EMG activities in fatigue duration. Journal of medical systems, 34(2), pp.131-138.

Response 1: We thank the reviewer for the insightful comment highlighting the absence of references to important studies on EEG-EMG signal coherence and its age dependence. We acknowledge the relevance of this feedback and have revised the manuscript to address this gap, with updates incorporated into Section 1. Introduction and Section 4. Conclusions to contextualize the neurophysiological basis of our EEG analysis and to the Discussion to propose future work to integrate EEG-EMG coherence metrics to improve dementia severity assessment. This revision strengthens the scientific foundation of our study and aligns it with established research. In the new version of the paper, major changes are highlighted in yellow.

Section 1. Introduction

“Moreover, the coherence of the cerebral EEG and muscular electromyogram (EMG) signal, the extent to which the activity of EEG-EMG can be synchronized is a significant method of studying sensorimotor integration and its variation during neurological disorders such as dementia. The studies cited by the reviewer form a good foundation to this study. James et al. have demonstrated that EEG-EMG coherence and cumulant changes over time and the corticospinal oscillations in the cortex reduce with ages in a client [20]. As Mima and Hallett have provided a detailed review of corticomuscular coherence, and its importance in motor control, this could explain the interpretation of EEG patterns [21]. Moreover, the physiological principles and clinical implications of our EEG-EMG frequency analysis have been elucidated by Grosse et al., and thus justify the relevance of using high frequency range in identifying motor-related aberrations [22]. Liu et al. have provided another evaluation of the corticomuscular coherence applications, but focus on its possible use in monitoring neurodegenerative diseases, which is consistent with our aims of dementia classification [23]. Tuncel et al. have explored the concept of time-frequency coherence in the state of tiredness and offer a perspective into the interactions between dynamic signals that can enhance our preprocessing resiliency [24].”

Section 4. Conclusions

“Moreover, the integration of EEG-EMG coherence metrics to improve dementia severity assessment.”

Comments 2: The introduction does not indicate the influence of age and psychiatric illness on the individual EEG frequency profile.

Response 2: We thank the reviewer for the insightful comment regarding the absence of discussion on the influence of age and psychiatric illness on the individual EEG frequency profile in the introduction. To address this, we have expanded Section 1. Introduction to include a paragraph discussing how age and psychiatric illness shape individual EEG frequency profiles. In the new version of the paper, major changes are highlighted in yellow.

“Additionally, age and psychopathic illness are known to significantly change the EEG frequency profiles, which are important for explaining brain activity in neurological and cognitive studies. The aging is associated with a change in EEG power, including a reduction in alpha (8-13 Hz) and an increase in Delta (1-4 Hz) and Thea (4-8 Hz) activity, reflecting converted cortical excitability and connection [18]. Similarly, psychiatry diseases, such as depression or anxiety, normal comoridities in dementia, can change the EEG profile by increasing the slow wave activity or by reducing beta (13-30 Hz) coherence [19].”

Comments 3: The significance of the obtained results is not presented taking into account the influence of the individual EEG frequency pattern (a decrease in the dominant frequency with age and depending on the psychiatric illness)

Response 3: We appreciate the reviewer’s valuable comment about how individual EEG frequency patterns, including the decrease in dominant frequency associated with age and psychiatric illness, affect the significance of our analysis of the results obtained. The influence of age and psychiatric illness on EEG frequency profiles, including a reduction in dominant frequency (alpha and beta bands) with aging and alterations due to conditions like epilepsy or dementia, has been thoroughly investigated in our previous work [cite previous study, e.g., Author et al., Year]. That study elucidated how these factors modulate EEG spectral characteristics and their implications for dementia classification, providing a critical foundation for the current research.

The primary aim of this study is to classify EEG-based recordings during a working memory task using the MobileNetV2 convolutional neural network (CNN) to assess dementia severity. Our focus is on identifying the most efficient feature extraction layer (convolutional layers ('Conv-1'), batch normalization ('Conv-1-bn'), clipped ReLU ('out-relu'), 2-D Global Average Pooling ('global-average-pooling2d1'), and fully connected ('Logits') layers) from the MobileNetV2 CNN architecture, rather than revisiting frequency pattern dependencies. MobileNetV2, as a CNN, is optimized to process two-dimensional spectrogram images derived from preprocessed EEG data (using conventional filters and discrete wavelet transformation), rather than directly utilizing frequency features or sequential data.

The significance of our results—demonstrating 95.9% and 95.3% accuracies with the global-average-pooling2d-1 and Logits layers, respectively—lies in validating the efficacy of deep processing algorithms for differentiating NC, MCogImp, and VasD patients.

Comments 4: The captions to the figures have not been corrected (too small and what the letters mean is not deciphered)

Response 4: We thank the reviewer for the valuable comment regarding the figure captions. We confirm that all suggested corrections have been implemented in the revised manuscript.  

3. Additional clarifications

We appreciate your guidance and look forward to your feedback on our revised submission.