PREreview del XAI-based Data Visualization in Multimodal Medical Data

QUESTION: What are the main strengths of the manuscript?

1. Comprehensive coverage: The manuscript provides an extensive review with ~300 references, demonstrating thorough research of the literature on XAI in multimodal medical data.

2. Practical examples and case studies: The paper includes engaging, tangible examples (e.g., prosthetic hand control systems, skin cancer diagnosis) that make abstract concepts accessible.

3. Educational value: Explains complex XAI methods (LIME, SHAP, Grad-CAM) in a digestible way for readers with basic AI background, bridging the gap between technical and clinical domains.

4. Structured taxonomy: Provides a helpful categorization of XAI techniques (perturbation, concept, and example-based methods) with clear explanations of their applications.

5. Resource compilation: Offers a comprehensive catalog of multimodal medical datasets used in XAI research, serving as a valuable reference.

6. Visual aids: Includes well-designed figures and schematic diagrams that enhance understanding of complex concepts.

QUESTION: What are the main weaknesses of the manuscript?

1. Unclear paper type: The title and abstract don’t clearly indicate this is a systematic review/survey, making it difficult for readers to discover and understand the paper’s purpose.

2. Excessive length: The manuscript is too long (60 pages) without adequate use of supplementary materials to offload figures and tables.

3. Poor section organization: Confusing subsection numbering (A, B within sections) and unclear structural flow, particularly in the introduction.

4. Missing methodological details: Lacks essential information about the systematic review process (number of papers screened, inclusion/exclusion criteria, search strategy, flowchart).

5. Inconsistent terminology: Uses both “survey” and “systematic review” inconsistently throughout the paper.

6. Limited reproducibility: No clear methodology section explaining how the review was conducted, making it difficult to validate or reproduce.

QUESTION: Use the space below to assemble your Summary paragraph pulling sentences from your notes in STEP 2 and your answers to the above questions.

This manuscript presents a comprehensive review of explainable AI (XAI) techniques for visualizing and interpreting multimodal medical data, focusing on how these methods can enhance clinical decision-making by making AI model reasoning transparent to healthcare professionals. The authors explore three main XAI strategies (perturbation, concept, and example-based methods) and their applications across diverse medical data types including imaging, electronic health records, and omics data. While the paper demonstrates extensive research with approximately 300 references and provides valuable practical examples that bridge the technical-clinical gap, it suffers from structural and methodological weaknesses. The manuscript’s length (60 pages) and unclear presentation as a systematic review, combined with missing methodological details about the review process, hinder its accessibility and reproducibility. Despite these limitations, the work offers significant value as an educational resource and comprehensive catalog of XAI applications in healthcare, particularly for researchers and practitioners seeking to understand how explainable AI can address the critical need for interpretable machine learning in medical settings.

QUESTION: Use the space below to assemble your Evidence and Example section, drawing from the list of major and minor concerns and the relative suggestions on how to address them from STEP 4.

Major Issues:

1. Title and Abstract Clarity: The title “XAI-based Data Visualization in Multimodal Medical Data” misleadingly suggests a methods paper rather than a systematic review. The abstract fails to use key terms like “review” or “survey,” making the paper difficult to discover through academic databases. Recommendation: Revise the title to include “A Systematic Review of” and explicitly state the review nature in the abstract’s opening sentence.

2. Missing Methodological Framework: The paper lacks essential systematic review methodology including search strategy, inclusion/exclusion criteria, and a PRISMA-style flowchart showing paper selection process. Recommendation: Add a dedicated methodology section (Section B could be expanded) with detailed search terms, databases used, screening process, and quantitative summary of papers reviewed.

3. Structural Organization: The current section numbering system (using A, B within numbered sections) creates confusion and impedes navigation through the 60-page document. Recommendation: Implement hierarchical numbering (e.g., 4.1, 4.2) and consider adding an index with hyperlinks to major sections.

Minor Issues:

1. Length Management: The manuscript could benefit from moving detailed tables and supplementary figures to appendices. Recommendation: Relocate comprehensive dataset tables and detailed technical diagrams to supplementary materials while retaining key summary tables in the main text.

2. Terminology Consistency: The paper inconsistently refers to itself as both a “survey” and “systematic review.” Recommendation: Choose one term and use it consistently throughout, including in the conclusion where both terms currently appear.

3. Reference to Methodology Standards: The review methodology lacks citation to established systematic review guidelines. Recommendation: Reference appropriate systematic review standards (e.g., PRISMA guidelines) and justify any deviations from standard practice.

QUESTION: Use the space below to write your Other Points section. Answering the following questions may guide you through writing this section.

Other Points:

Positive Aspects:

·      The manuscript serves as an excellent educational resource, particularly for readers with basic AI backgrounds who need to understand XAI applications in healthcare

·      The practical examples (prosthetic control systems, skin cancer diagnosis) effectively demonstrate real-world applications

·      The comprehensive dataset compilation provides valuable reference material for future researchers

·      The explanation of local vs. global explainability concepts is particularly well-executed

Suggestions for Enhancement:

·      Consider developing an online companion resource or interactive tool to help readers navigate the extensive content

·      The extensive reference list could be better organized thematically to aid readers seeking specific types of studies

·      Readers might benefit from inclusion of a quantitative analysis of the reviewed literature (e.g., trends over time)

Target Audience Considerations: While the paper aims to bridge the gap between AI developers and medical practitioners, its current length and technical depth may limit accessibility to clinical audiences. The authors might consider developing a shorter, more clinically-focused companion piece highlighting key takeaways for healthcare professionals.

Publication Readiness: Despite the identified issues, the manuscript represents a significant contribution to the field and merits publication after addressing the major structural and methodological concerns. The comprehensive nature of the review and the quality of the technical content outweigh the presentation issues, which are correctable through revision.

QUESTION: What one thing from this work have you learned?

Summary of Key Learning Points:

The reviewers gained valuable insights into both technical XAI methods and conceptual frameworks:

1. Technical Understanding: The manuscript effectively explains complex XAI models (LIME, SHAP) and their practical applications in an accessible way for readers with basic AI backgrounds.

2. Conceptual Framework: The distinction between local and global explainability emerged as a key learning point - understanding how these two approaches work together to provide layered understanding of AI decision-making in medical contexts.

3. Practical Application: The skin cancer diagnosis example (p17) particularly helped illustrate how local explanations reveal why a model made specific predictions for individual patients, while global explanations show overall model behavior patterns across entire patient populations.

QUESTION: Any final positive remarks?

Summary of Final Positive Remarks:

The reviewers highlighted several commendable aspects of the manuscript:

1. Research Quality: The paper demonstrates extensive research that has been expertly condensed while maintaining accessibility through engaging examples.

2. Broader Impact Awareness: The manuscript effectively illustrates the importance of XAI tools not just for technical users, but for all stakeholders affected by AI decisions in healthcare, serving as a reminder that research should ultimately serve people.

3. Visual Communication: The figures significantly enhance understanding of complex concepts, making the technical content more accessible to readers.

QUESTION: Would you recommend this manuscript to others to read?

Summary of Recommendations:

All reviewers would recommend this manuscript to others, but with important caveats:

·      Target audience: Highly recommended for AI researchers and practitioners working specifically in the medical domain, particularly those interested in explainable AI (XAI) applications

·      Length concern: The comprehensive nature (60 pages) makes it excellent for specialized readers but potentially discouraging for general audiences

·      Value proposition: Despite its length, the manuscript’s thoroughness and comprehensive coverage make it a valuable resource for those working in this specific field

QUESTION: Would you recommend this manuscript for journal publication?

Yes, after the major issues have been addressed this article can be a very useful reference for the intended audience.

Competing interests

The authors declare that they have no competing interests.