Artificial Intelligence and the Expanding Universe of Cardio-Oncology: Beyond Detection Toward Prediction and Prevention of Therapy-Related Cardiotoxicity—A Comprehensive Review

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The review argues that cardiotoxicity (heart damage from cancer treatments like chemotherapy, targeted therapies, immunotherapy, and radiation) is a major problem for cancer survivors. As cancer survival improves, cardiovascular complications have become a leading cause of morbidity/mortality in this population.

There are the following contributions 

Why Artificial Intelligence in Cardio-Oncology?

• AI integrates biomarkers, exosomes, imaging (radiomics/dosiomics), ECG, genomics/proteomics.
• Reduces subjectivity & inter-observer variability.
• Distinguishes classical ML (feature-based, interpretable: logistic regression, random forest, SVM, gradient boosting) vs Deep Learning (automatic feature learning: CNNs for images, good for echo strain, CMR mapping).
• Discusses supervised/unsupervised/reinforcement learning.
• Importance of validation (AUC, calibration) and standards: TRIPOD-AI, PROBAST-AI, CLAIM

But: Weak evidence and lack of validation (the biggest problem in the field, repeated in almost all articles).
Real-world implementation barriers (bias, access inequality, ethics, integration) are not explored in sufficient depth. Less emphasis on new technologies (AI-ECG, wearable, multimodal LLMs).

Author Response

Response:

We sincerely thank the reviewer for the thorough, thoughtful, and overall positive evaluation of our work. We particularly appreciate the clear and accurate summary of the main aims and contributions of the manuscript, including the clinical relevance of cardiotoxicity in cancer survivors and the potential role of artificial intelligence in cardio-oncology. 

We thank the reviewer for highlighting the importance of real-world implementation barriers and emerging technologies. We fully agree that these aspects are critical for the clinical translation of AI in cardio-oncology. In response, we have substantially expanded and restructured the “Limitations and Challenges” section to explicitly address (i) data quality and dataset bias, (ii) access inequality and algorithmic bias, (iii) infrastructural and integration barriers (including ECG data availability and vendor lock-in), (iv) regulatory and ethical considerations, and (v) the clinical risks related to opacity, lack of interpretability, uncertainty estimation, and failure-mode analysis. We now explicitly discuss how these factors limit trust, safety, and real-world deployment in high-stakes cardio-oncology decision-making. We also introduced comparative summary tables that synthesize key characteristics of the reviewed studies, highlighting for each domain the clinical task, AI methodology, study population, validation strategy, explainability of the results and reported performance, in order to clarify even more the real world barriers.

Reviewer 2 Report

Comments and Suggestions for Authors

This review presents an overview of artificial intelligence applications in cardio-oncology, with extensive coverage across imaging modalities, ECG analysis, biomarkers, and radiotherapy dose modeling. However, in its current form, the paper reads more as an annotated literature compilation than a methodologically rigorous, analytically critical review. Several structural, conceptual, and methodological issues limit its scholarly depth and clinical translational value.

Most critically, the manuscript does not address explainability, interpretability, or trustworthiness of AI models, despite relying heavily on deep learning systems that are intrinsically opaque.

The manuscript extensively discusses deep learning models but fails to address how these models are interpreted, validated clinically, or trusted by clinicians. There is no systematic discussion of: Saliency maps,  Feature attribution methods, Attention visualization, Uncertainty estimation, Failure mode analysis. This is a major conceptual gap, especially in a clinical domain where AI outputs may influence treatment interruption or continuation. The authors must introduce a dedicated subsection on Explainable AI (XAI), explicitly referencing and synthesizing insights from: A Survey on Explainable Artificial Intelligence (XAI) Techniques for Visualizing Deep Learning Models in Medical Imaging. This survey should be used to Categorize explainability methods (post-hoc vs intrinsic), Discuss their relevance to cardio-oncology imaging. Highlight modality-specific XAI practices, Critically assess whether cited studies used any explainability mechanisms.

Large portions of Sections 4 and 5 summarize individual studies sequentially, with limited synthesis.

Introduce comparative summary tables (e.g., modality × task × AI method × validation × explainability).

Despite being a review, the manuscript does not clearly specify research strategy.

The manuscript repeatedly reports AUC, Dice scores, ICC, and MAE but does not critically assess:

Duplicate headings (e.g., “4.1. Echocardiography” appears multiple times)

Excessive length in some modality subsections without synthesis

Lack of a limitations-of-AI section beyond dataset size

Comments on the Quality of English Language

Minor grammatical inconsistencies.

Author Response

• “This review presents an overview of artificial intelligence applications in cardio-oncology, with extensive coverage across imaging modalities, ECG analysis, biomarkers, and radiotherapy dose modeling. However, in its current form, the paper reads more as an annotated literature compilation than a methodologically rigorous, analytically critical review. Several structural, conceptual, and methodological issues limit its scholarly depth and clinical translational value.

Most critically, the manuscript does not address explainability, interpretability, or trustworthiness of AI models, despite relying heavily on deep learning systems that are intrinsically opaque.

The manuscript extensively discusses deep learning models but fails to address how these models are interpreted, validated clinically, or trusted by clinicians. There is no systematic discussion of: Saliency maps,  Feature attribution methods, Attention visualization, Uncertainty estimation, Failure mode analysis. This is a major conceptual gap, especially in a clinical domain where AI outputs may influence treatment interruption or continuation. The authors must introduce a dedicated subsection on Explainable AI (XAI), explicitly referencing and synthesizing insights from: A Survey on Explainable Artificial Intelligence (XAI) Techniques for Visualizing Deep Learning Models in Medical Imaging. This survey should be used to Categorize explainability methods (post-hoc vs intrinsic), Discuss their relevance to cardio-oncology imaging. Highlight modality-specific XAI practices, Critically assess whether cited studies used any explainability mechanisms.”

Response:

We thank the reviewer for this thoughtful and detailed critique and fully agree that explainability, interpretability, and trustworthiness are critical issues for the clinical deployment of AI systems, particularly in a high-stakes domain such as cardio-oncology.

We also agree that these topics deserve explicit discussion, and in response to this comment, we have strengthened the manuscript by expanding the discussion on interpretability, clinical trust, and transparency of AI models, with particular emphasis on their relevance to clinical decision-making.

However, we would like to clarify the intended scope and target audience of this review. This manuscript was conceived as a clinically oriented narrative review, primarily aimed at cardiologists, oncologists, and clinicians involved in the care of cancer patients. Our goal was to provide a comprehensive, accessible overview of how AI is being applied in cardio-oncology, its potential clinical value, and its current limitations, rather than to offer a technically exhaustive explanation of AI methodology.

We carefully reviewed the suggested survey article (“A Survey on Explainable Artificial Intelligence (XAI) Techniques for Visualizing Deep Learning Models in Medical Imaging”). While this is an excellent and technically rigorous contribution, it is written from a computer science perspective and addresses concepts (e.g., detailed categorizations of attribution methods, architectural interpretability, and visualization taxonomies) that go well beyond the level of technical depth typically familiar to most clinicians. Incorporating this material in a comprehensive way would substantially shift the focus of the manuscript away from its clinical and translational intent and risks making the review less accessible.

That said, we fully agree with the reviewer that the clinical implications of model opacity, trust, and explainability must be explicitly addressed. We have therefore revised the manuscript to:
(i) introduce a dedicated subsection discussing interpretability, explainability, and clinical trust in AI;
(ii) explain, in clinician-oriented terms, why “black-box” behavior is problematic in cardio-oncology decision-making;
(iii) outline, at a conceptual level, the main families of explainability approaches (e.g., post-hoc vs intrinsic methods); and
(iv) critically note that most studies in the current cardio-oncology literature still lack robust or systematic use of explainability, uncertainty estimation, or failure-mode analysis.

We believe this approach preserves the clinical focus and readability of the review while directly addressing the important conceptual gap highlighted by the reviewer and reinforcing the translational challenges that must be overcome before AI tools can be safely and confidently adopted in routine care.

We added the following paragraphs in the “Limitations and challenges” section:

“Despite the rapid expansion of artificial intelligence applications in cardio-oncology, the clinical adoption of these tools remains strongly influenced by issues of interpretability, explainability, and trust. Many of the most powerful models currently used in medical imaging, electrocardiography, and multimodal data integration rely on deep learning architectures that function largely as “black boxes,” meaning that the internal reasoning leading to a given prediction is not directly accessible to the clinician .

In cardio-oncology, this limitation is particularly important because AI outputs may influence high-stakes decisions, such as whether to interrupt, modify, or continue potentially life-saving cancer therapies, or whether to initiate cardioprotective treatment. In such contexts, clinicians must not only know what the model predicts, but also have a reasonable understanding of why a given patient is classified as high or low risk.

From a conceptual perspective, explainability methods can be broadly divided into two main categories. Intrinsic (or interpretable-by-design) models are constructed in a way that their decision process is inherently transparent, as is often the case with simpler machine learning models based on a limited number of clinically meaningful variables. In contrast, post-hoc explainability methods attempt to provide explanations after a complex model has generated its output, for example by highlighting image regions, signal segments, or input features that contributed most strongly to a given prediction. While such approaches can improve human understanding of model behavior, they do not fully eliminate the fundamental opacity of deep learning systems.

In cardiovascular imaging and ECG analysis, post-hoc visualization techniques are increasingly used to illustrate which regions of an image or which portions of a signal drive a model’s prediction. When these highlighted areas correspond to clinically plausible structures, such as the myocardium, valves, or specific ECG segments, they may increase clinician confidence and help detect spurious correlations. However, these methods remain indirect and do not guarantee that the model’s reasoning is physiologically sound or robust across populations.

Another closely related issue is the limited assessment of uncertainty and failure modes in most current studies. Very few published models in cardio-oncology explicitly report prediction uncertainty, analyze cases of systematic failure, or define situations in which the algorithm should not be trusted. This is problematic, as overconfident but incorrect predictions may lead to inappropriate reassurance or unnecessary treatment modifications.”

 

• “Large portions of Sections 4 and 5 summarize individual studies sequentially, with limited synthesis.”

Response:

We thank the reviewer for this insightful comment. We acknowledge that Sections 4 and 5 present a sequential discussion of individual studies. This approach was intentional and aligned with the primary aim of the manuscript, which was designed as a narrative and comprehensive review rather than a systematic or meta-analytic synthesis.

Given the heterogeneity of study designs, patient populations, cancer therapies, AI methodologies, and outcome definitions in the current cardio-oncology literature, our objective was to provide clinicians and researchers with a broad, structured overview of the available evidence. By summarizing individual studies in detail, we aimed to offer a practical reference for clinicians caring for oncology patients, as well as for cardio-oncology specialists seeking to understand how different AI approaches have been applied across diverse clinical contexts.

We agree that synthesis is essential for interpretability, and where possible, we have emphasized overarching themes, recurring methodological limitations, and common performance trends across studies. 

We believe that this comprehensive, study-by-study narrative approach ultimately supports the manuscript’s goal of informing clinical practice and guiding future research by clearly illustrating both the promise and the limitations of AI applications in cardio-oncology.

• “Introduce comparative summary tables (e.g., modality × task × AI method × validation × explainability). “

Response:

We thank the reviewer for this valuable and constructive suggestion. In response, we have introduced comparative summary tables that synthesize key characteristics of the reviewed studies.

We also added the following explanatory paragraph: “For clarity and synthesis, the main studies discussed in this review are summarized in modality-specific comparative tables covering echocardiography (Table 1), cardiac magnetic resonance (Table 2), computed tomography and radiomics/dosiomics (Table 3), nuclear medicine imaging (Table 4), electrocardiography (Table 5), and multimodal/translational approaches including biomarkers, genomics, proteomics, and extracellular vesicles (Table 6), highlighting for each domain the clinical task, AI methodology, study population, validation strategy, and reported performance.“

• “Despite being a review, the manuscript does not clearly specify research strategy.”

Response:

We thank the reviewer for this important observation. We acknowledge that the manuscript did not initially describe the research strategy in sufficient detail. As this work was conceived as a narrative review, our intent was to provide a comprehensive and clinically oriented overview of the rapidly evolving field of artificial intelligence in cardio-oncology, rather than to conduct a systematic review or meta-analysis.

In response to this comment, we have clarified the literature identification and selection approach in the Methods section, including the databases consulted, key search terms, time frame, and general inclusion criteria used to identify relevant studies. While we did not apply formal systematic-review methodologies (e.g., PRISMA), we aimed to ensure broad coverage of high-impact and clinically relevant publications across modalities and AI approaches.

We believe that this clarification improves transparency and allows readers to better interpret the scope and intent of the review, while remaining consistent with the narrative nature of the manuscript.

Hereby is the new “Materials and methods” section that was added to the manuscript:

“4. Materials and methods

4.1. Review Design

This manuscript was designed as a comprehensive narrative review aimed at providing a clinically oriented overview of current and emerging applications of artificial intelligence in cardio-oncology, with a focus on cancer therapy–related cardiotoxicity. Given the marked heterogeneity of study designs, populations, AI methodologies, data modalities, and outcome definitions in this rapidly evolving field, a formal systematic review or meta-analysis was not considered methodologically appropriate. Instead, a structured narrative approach was adopted to ensure broad coverage, clinical interpretability, and translational relevance.

4.2. Literature Search Strategy

A comprehensive literature search was conducted in the following electronic databases: PubMed/MEDLINE, Scopus, and Web of Science.

The search strategy incorporated relevant keywords and combinations thereof, including: cardio-oncology; cancer therapy–related cardiotoxicity; artificial intelligence; machine learning; deep learning; echocardiography; cardiac magnetic resonance; computed tomography; nuclear imaging; electrocardiography; radiomics; dosiomics; biomarkers; proteomics; genomics; extracellular vesicles.

Evidence was identified focusing on applications in electrocardiography, biomarkers, proteomics, extracellular vesicles, genomics, advanced imaging (echocardiography, cardiac magnetic resonance, computed tomography, and nuclear imaging), and radiotherapy dose modeling (dosiomics). Translational insights from animal models and in vitro systems were also included. In addition, the reference lists of relevant reviews and key original articles were manually screened to identify further pertinent studies.

4.3. Inclusion Criteria

Studies were considered eligible for inclusion if they met the following criteria:

1. Investigated applications of artificial intelligence, machine learning, or deep learning in the prediction, detection, monitoring, or prevention of cancer therapy–related cardiotoxicity or cardiovascular complications in oncology.
2. Included any study type, including clinical, translational, preclinical, in silico, or in vitro studies.
3. Addressed at least one of the following domains: cardiovascular imaging, electrocardiography, biomarkers, multi-omics, extracellular vesicles, or radiotherapy dose modeling.
4. Presented original research or methodologically relevant translational studies.
5. Used artificial intelligence for prediction, classification, segmentation, risk stratification, data integration, or decision support in a cardio-oncology–relevant context.
6. Were published in peer-reviewed journals.
7. No strict restrictions were applied regarding publication date in order to ensure a broad and comprehensive overview of the field.

4.4. Exclusion Criteria

Studies were excluded based on the following criteria:

1. Duplicate publications or overlapping reports from the same study population, in which case the most complete or most recent version was retained.
2. Articles whose title and abstract did not indicate a clear focus on artificial intelligence applications in cardio-oncology or cancer therapy–related cardiovascular toxicity.
3. Papers discussing artificial intelligence in oncology or cardiology without relevance to cardiotoxicity or cardiovascular complications of cancer therapy.
4. Studies unrelated to medical or biomedical applications.
5. Patents, book chapters, editorials, commentaries, opinion pieces, and non–peer-reviewed literature.

4.5. Data Extraction and Synthesis

From each selected study, we extracted information on the clinical context, data modality, artificial intelligence methodology, clinical task, validation strategy, reported performance metrics, and main limitations. The evidence was synthesized thematically and by modality, emphasizing clinical relevance, translational potential, and methodological trends rather than quantitative pooling of results.

4.6. Methodological Quality Considerations

Given the variable methodological quality and reporting standards in the artificial intelligence literature, studies were critically appraised with reference to TRIPOD-AI, PROBAST-AI, and CLAIM guidelines. These frameworks were used to contextualize common limitations, including small cohort sizes, risk of bias, lack of external validation, incomplete reporting, and limited assessment of calibration and clinical utility, rather than to formally exclude studies.

4.7. Rationale for Narrative Approach

This field is characterized by rapid technological evolution, heterogeneous methodologies, and a predominance of exploratory and proof-of-concept studies. Under these circumstances, a narrative synthesis was considered more appropriate than a systematic review or meta-analysis, allowing a broader, clinically meaningful, and integrative perspective while explicitly highlighting both opportunities and current limitations.”

 

 

• “The manuscript repeatedly reports AUC, Dice scores, ICC, and MAE but does not critically assess:

Duplicate headings (e.g., “4.1. Echocardiography” appears multiple times)”

Response:

We also thank the reviewer for identifying the issue of duplicate section headings. This was an oversight, and the manuscript structure has been revised to eliminate redundant headings and improve clarity and readability.

• “Excessive length in some modality subsections without synthesis”

Response:

We thank the reviewer for this comment. We acknowledge that some modality-specific subsections are relatively lengthy. This reflects the intentional design of the manuscript as a comprehensive narrative review, aimed at providing clinicians and researchers with a detailed overview of the diverse AI applications across imaging, ECG, biomarkers, and multimodal data in cardio-oncology.

Given the marked heterogeneity in methodologies, patient populations, clinical endpoints, and validation strategies, we elected to present individual studies in sufficient detail to preserve clinical and methodological context. 

We believe this balanced approach maintains the educational value of a comprehensive narrative review while addressing the reviewer’s concern regarding synthesis.

• “Lack of a limitations-of-AI section beyond dataset size”

Response:

We thank the reviewer for this important comment. We respectfully note that the manuscript includes a dedicated subsection addressing limitations and challenges of AI in cardio-oncology that extend well beyond dataset size.

Specifically, this section discusses multiple, interrelated limitations, including:
(i) data quality, annotation errors, and embedded biases;
(ii) substantial heterogeneity across study designs, patient populations, cancer types, cardiovascular risk profiles, definitions of cardiotoxicity, biomarkers, and imaging modalities;
(iii) small or selected cohorts and the widespread lack of external validation;
(iv) infrastructural and interoperability barriers, particularly the limited availability of raw digital ECG data and vendor-related constraints;
(v) challenges related to individualized pathophysiology in cancer therapy–related cardiovascular risk prediction;
(vi) regulatory, ethical, and governance considerations, including data privacy, accountability, and patient safety;
(vii) limited model interpretability and the “black box” nature of many AI systems;
(viii) the risk of algorithmic bias and reinforcement of health disparities due to underrepresentation of minority populations; and
(ix) publication bias, which may contribute to overly optimistic assessments of AI performance.

We hope this clarification highlights that the manuscript provides a balanced and critical appraisal of AI limitations in cardio-oncology, consistent with the reviewer’s suggestion.

Reviewer 3 Report

Comments and Suggestions for Authors

In the index report, the authors explore critical and emerging area in stem cell research by exploring the applications of artificial intelligence (AI) in the field of cardio-oncology. It highlights how AI is transforming the detection, prediction, and prevention of cancer therapy-related cardiotoxicity. Interestingly, the review highlights the unique insights that can be gained from applying AI to electrocardiography (ECG) data. AI-enabled ECG analysis has demonstrated the ability to detect subtle changes in cardiac function, even before traditional measures like ejection fraction show any detectable changes. This early identification of subclinical cardiac dysfunction holds great promise for proactive management and prevention of cardiotoxicity.The abstract is well written and adequality captures the salient takeaways from their research. overall, the paper is well written and largely free of grammatical errors. While the article provides valuable insights into the potential pathophysiological mechanisms, diagnostic challenges, and treatment considerations, there are a few potential weaknesses and areas for further exploration:-

1. The manuscript highlights the substantial heterogeneity across existing studies in terms of study designs, patient populations, cancer types, cardiovascular risk profiles, definitions of treatment-related cardiotoxicity, biomarker assays, and imaging techniques. This introduces potential selection and reporting biases, limits the generalizability of findings, and underscores the need for standardized, multicenter studies with larger, more diverse populations to develop broadly applicable clinical guidelines
2. Many of the studies discussed have small or selected patient cohorts, which limits the statistical power and generalizability of the findings.
3. The authors note that a key limitation of the current literature is the lack of external validation for many of the proposed AI models and algorithms. Without rigorous external validation, the reliability and applicability of these tools in real-world clinical practice remains uncertain.
4. The author acknowledges the potential for AI systems to unintentionally reinforce healthcare disparities and inequities when structural biases are present in the training data. This is an important consideration that needs to be addressed to ensure fair and equitable implementation of AI tools in cardio-oncology.
5. The authors caution that publication bias may affect the current literature, as studies reporting positive results are more likely to be published than those with negative or inconclusive findings. This can lead to an overly optimistic perception of AI performance.
6. Add 2-3 representative figures and tables for the readership. Facilitates understanding and interests of the readers

Overall, the review provides a balanced and critical assessment of the current state of AI in cardio-oncology, highlighting key limitations and areas for improvement to ensure the reliable and equitable development and implementation of these technologies.

Author Response

We sincerely thank the reviewer for their thoughtful, detailed, and constructive evaluation of our manuscript. We greatly appreciate the reviewer’s accurate summary of the scope and objectives of our review. We acknowledge and agree with the reviewer’s important observations. Finally, we thank the reviewer for the helpful suggestion to include representative figures and tables. In response, we have added illustrative figures/tables to improve clarity, enhance reader engagement, and facilitate understanding of key concepts and workflows discussed in the review.

 

Round 2

Reviewer 2 Report

Comments and Suggestions for Authors

Author has made all the suggested updates.

Reviewer 3 Report

Comments and Suggestions for Authors

the authors have satisfactorily responded to most concerns raised during the recent review .