Machine Learning and Deep Learning Approaches in Lifespan Brain Age Prediction: A Comprehensive Review

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

I have really enjoyed reading the submitted review. It resumes more than 50 very recent studies, where machine and deep learning tools are used to predict patients' brain age through neuroimaging analysis.

The topic is of interest, as demonstrated by the number of related works published in the last few years involving many different approaches.

The paper is well organized and written; it also overviews many neural networks with a technical yet perfectly comprehensible report.

In the following are my remarks and suggestions to improve the manuscript.

1) I think the paper misses a chapter focusing on the datasets. ML and DL approaches strongly depend on the quality and, even before, on the kind of data they work on. The authors always talk about "neuroimaging data" or, in general, "data" and "comprehensive information" (line 247). It is not clear if (and when) the cited data sets contain images or also patients' metadata. How do ML tools work on images? Are there 2D or 3D brain images? I think Section 4 (or Tables 1 and 2) should include specifications about the kind of analyzed data.

2) In Tables 1 and 2, what does the R column refer to? A better description of the R meaning and role could also be beneficial for comments on Table 4.
In addition, how are the MSE values computed? Have you run the proposed algorithms on the entire dataset to compute comparable metrics? Sometimes the R-value is missing in Table 2: why?

3) In Figure 4, I do not understand the meaning of the lines connecting couples of boxplots. I think that a brief and concise explanation of the U test could be useful.

4) Again, on the boxplots in Figures 4 and 5, what does each data point represent? One paper (using one of the considered models or networks)?

5) The conclusions should be expanded. After reading the review, many questions arise.
- Are there features (on metadata or images) that the literature leverages as the most important ones? 
- Are there frequent issues or difficulties in making brain-age predictions? Which ones?
- Which are the future research directions that have been mainly suggested by the reviewed papers?

6) One curiosity: is explainable AI considered in the papers? You have never talked about prediction interpretability and explainability.

7) A few typos should be fixed. For instance, in line 381, a capital "T" is missing. I also think a few notations must be introduced, such as LR (line 133) and AD (line 522).

Comments on the Quality of English Language

The paper is well written in fluent and correct English.

Author Response

Dear Editors and Reviewers,

Firstly, we would like to express our sincere gratitude for the time and effort you have dedicated to reviewing our manuscript. Your insightful comments have been instrumental in enhancing the quality of our paper.

In accordance with your valuable feedback, we have meticulously revised our manuscript. Below, we address each of the comments provided by the reviewers.

Thank you once again for your thorough review and valuable feedback. and we look forward to your continued guidance throughout the review process.

Sincerely,

Lan Lin

Beijing University of Technology

I have really enjoyed reading the submitted review. It resumes more than 50 very recent studies, where machine and deep learning tools are used to predict patients' brain age through neuroimaging analysis.

The topic is of interest, as demonstrated by the number of related works published in the last few years involving many different approaches.

The paper is well organized and written; it also overviews many neural networks with a technical yet perfectly comprehensible report.

In the following are my remarks and suggestions to improve the manuscript.

Q1: I think the paper misses a chapter focusing on the datasets. ML and DL approaches strongly depend on the quality and, even before, on the kind of data they work on. The authors always talk about "neuroimaging data" or, in general, "data" and "comprehensive information" (line 247). It is not clear if (and when) the cited data sets contain images or also patients' metadata. How do ML tools work on images? Are there 2D or 3D brain images? I think Section 4 (or Tables 1 and 2) should include specifications about the kind of analyzed data.

Thank you very much for your valuable suggestions. Indeed, ML and DL methods are highly dependent on the quality of the data. We apologize for any unclear expressions in the original manuscript. In the latest revised version, we have reorganized Section 2 to include a foundational description of the data and to outline how neuroimaging data are processed from raw data into formats suitable for DL and ML. Additionally, we have included relevant information about data modality in Tables 1 and 2.

Q2: In Tables 1 and 2, what does the R column refer to? A better description of the R meaning and role could also be beneficial for comments on Table 4.
In addition, how are the MSE values computed? Have you run the proposed algorithms on the entire dataset to compute comparable metrics? Sometimes the R-value is missing in Table 2: why?

In response to your suggestions, we have added a section in the reorganized Section 2 descript the performance evaluation criteria for brain age prediction models. This section provides a detailed explanation of the evaluation metrics used for brain age prediction models, including the meaning and role of R-values, the calculation formula for MAE (it seems you might have been referring to MAE in your question, as our manuscript does not report MSE performance), and the calculation formula for R-values. The absence of R-values in Tables 1 and 2 is because these articles do not utilize this metric. We have included a footnote in Tables 1 and 2 explaining the reason for the omission of R-values.

Q3: In Figure 4, I do not understand the meaning of the lines connecting couples of boxplots. I think that a brief and concise explanation of the U test could be useful.

We apologize for the lack of detailed description regarding the meaning of the connecting lines in Figure 4. In the latest revised version, we have provided a concise explanation of the U test calculation process in the paragraph describing Figure 4.

Q4: Again, on the boxplots in Figures 4 and 5, what does each data point represent? One paper (using one of the considered models or networks)?

In accordance with your suggestion, we have provided a detailed explanation of what the data points represent in the descriptive paragraphs for Figures 4 and 5.

Q5: The conclusions should be expanded. After reading the review, many questions arise.
- Are there features (on metadata or images) that the literature leverages as the most important ones? 
- Are there frequent issues or difficulties in making brain-age predictions? Which ones?
- Which are the future research directions that have been mainly suggested by the reviewed papers?

Given that this study encompasses participants across the full age spectrum, the most significant features often manifest differently across various age groups. This variation is discussed in detail in Section 5.1, within the interpretability-focused segment on AI.

In the latest revised version, we have summarized the current challenges and future directions in Section 5.3.

Q6: One curiosity: is explainable AI considered in the papers? You have never talked about prediction interpretability and explainability.

This is an excellent question. Indeed, the articles we reviewed do mention aspects related to model interpretability. In the latest revised version, we have added a discussion on model interpretability n Section 5.1.

Q7: A few typos should be fixed. For instance, in line 381, a capital "T" is missing. I also think a few notations must be introduced, such as LR (line 133) and AD (line 522).

Thank you very much for pointing out the errors in the manuscript. In the latest revised version, we have corrected the errors you mentioned and have thoroughly reviewed the text to avoid any other potential issues.

Reviewer 2 Report

Comments and Suggestions for Authors

Comments for author File: Comments.pdf

Author Response

Dear Editors and Reviewers,

Firstly, we would like to express our sincere gratitude for the time and effort you have dedicated to reviewing our manuscript. Your insightful comments have been instrumental in enhancing the quality of our paper.

In accordance with your valuable feedback, we have meticulously revised our manuscript. Below, we address each of the comments provided by the reviewers.

Thank you once again for your thorough review and valuable feedback. and we look forward to your continued guidance throughout the review process.

Sincerely,

Lan Lin

Beijing University of Technology

The review covers 52 peer-reviewed studies, assessing various ML and DL model architectures, their effectiveness, and the challenges in achieving precise brain age prediction across different age groups. The paper also discusses the strengths and limitations of these models, providing a synthesis of current state-of-the-art methods and guiding future research to improve early intervention strategies for neurodegenerative diseases. However, need some improvements:

Q1: Future reviews should emphasize the need for standardizing neuroimaging protocols and consider the development of harmonization techniques to address variability in datasets.

In clinical practice, neuroimaging data are often collected from different devices and under varying protocols, leading to significant variability in the datasets. Consequently, one challenge for brain age prediction models is to adapt to data originating from diverse protocols and equipment, thereby enhancing the model's generalizability in real-world applications. Current research indicates that deep learning models have demonstrated promising performance in this regard, showing good predictive capabilities with unseen data.

Q2: Please include a detailed analysis of computational costs, resource requirements, and practical feasibility of implementing DL models in clinical settings.

Thank you for your valuable suggestions regarding the computational cost, resource requirements, and practical feasibility of implementing deep learning models in clinical environments. While these are considerations for brain age prediction models, the issues related to computational cost and complexity are becoming less significant due to advancements in GPU technology. Training times can vary from a few hours to several weeks, depending on data volume, model architecture, and computational resources. However, once trained, clinical deployment of these models typically allows for results to be obtained within milliseconds.

Q3: Emphasizing the need for external validation across diverse populations and geographic regions can help ensure that models are widely applicable.

In response to your suggestion, we have emphasized in Section 5.2 the need for external validation across diverse populations and geographic regions can help ensure that models are widely applicable.

Q4: Please include some section that describing metrics that used in the current area. Please cite this paper as example of DL using in area of medicine: Muksimova, S., Umirzakova, S., Mardieva, S. and Cho, Y.I., 2023. Enhancing medical image denoising with innovative teacher student model-based approaches for precision diagnostics. Sensors, 23(23), p.9502.

Thank you very much for your suggestions. We have reorganized Section 2 and added relevant descriptions about neuroimaging data and electrophysiological source signals. In these descriptions, we discuss the current data preprocessing workflows in the field and cite the paper you recommended.

Q5: Please include section about Challenges in the current area.

Your suggestions prompted us to write the Section 5.3. In the latest revised version, we outline the current challenges and opportunities in the field, aiming to provide direction for researchers involved in this area.

Round 2

Reviewer 2 Report

Comments and Suggestions for Authors

Well done.