Skip to Content
MDPIMDPI

Advanced

by
  • Ali Gates 1,
  • Asif Ali 2 and
  • Patrick Dunn 1,*
  • et al.

Reviewer 1: Lei Hang Reviewer 2: Anonymous

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

After reviewing the document titled " Exploring an AI First Healthcare System," here are detailed section-wise comments and suggestions for improvement:

  1. Title / author affiliation formatting

Affiliations and superscripts are inconsistent/fragmented (line breaks and stray numbers). Journal metadata style should be consistent (author names, numbered affiliations, corresponding author email).

  1. Abstract

The Abstract presently asserts broad transformative impacts without noting constraints (validation needs, generalizability, regulatory hurdles). These risks overstating readiness for clinical deployment.

  1. Hypertension call-out (lines ~139–144).

Text suggests cuffless BP as a mature, reliable solution. Cuffless methods (photoplethysmography, pulse transit time) show promise but require calibration, are sensitive to motion/artifacts, and often lack robust validation across skin tones/conditions.

  1. Inpatient / surgical section (lines ~52–56; 183–189).

The draft conflates AI assistance with autonomous surgery. Regulatory and safety landscape differs drastically for assistive guidance vs autonomous actions.

  1. Architecture / ambient listening (lines ~308–313).

Ambient listening is described without discussion of consent, HIPAA/GDPR, data minimization, encryption, or legal/ethical governance.

  1. Bias/equity discussion (lines ~341–348).

The manuscript highlights bias risk but does not recommend practical mitigation strategies or metrics.

  1. Explainability/trust section (lines ~349–356).

“Explainable AI” is mentioned generically; the manuscript should distinguish interpretable models (e.g., logistic regression, decision trees) vs post-hoc explainers (SHAP, LIME) and limitations of each.

  1. Multiple sections discuss validation generally (e.g., Conclusion, Research agenda lines ~444–452).

Lack of specifics about how to validate AI in healthcare.

  1. Architecture & interoperability (lines ~300–307, ~340).

Interoperability is noted but missing concrete standards to guide implementation.

  1. Data privacy/security discussion (lines ~356–363).

Security is mentioned superficially; AI systems bring attack vectors (model inversion, poisoning, adversarial attacks).

  1. Research & Implementation Agenda / Conclusion (end of manuscript).

No statement about sharing code, models, or data (or reasons if not possible). For reproducibility, funders/journals increasingly require this.

  1. Typos, formatting, and small language fixes
  • “Arcutecture on an AI first healthcare system.” → “Architecture of an AI-first healthcare system.” (Figure 3 caption).
  • “the patients entire medical record” → “the patient’s entire medical record” (add possessive).
  • “post-visit … appointment follow up procedure” → hyphenate/consistency: “post-visit follow-up”.
  • Ensure consistent hyphenation of “AI-first”, “hospital-at-home”, and “cuffless”
  • Check references for consistent DOI format and capitalization.
  • “Key terminology in AI in healthcare.”->” Key technology in AI in healthcare.” (Figure 2 caption)
  1. Figures

The resolutions of some figure such as Figure 3 can be improved as the text inside is blurry. Moreover, the citations may be required as some of the figures may come from sources of Internet.

 

 

 

 

 

Author Response

We thank both reviewers for their careful reading of the manuscript and their constructive, detailed feedback. We have substantially revised the manuscript to address all comments. Below, we respond point by point to Reviewer 1 and Reviewer 2, indicating how each concern was addressed in the revised version.

Reviewer 1

Comment 1: Title / author affiliation formatting

Affiliations and superscripts are inconsistent; journal metadata style should be consistent.

Response:
We have corrected formatting inconsistencies in the title page and author affiliations. Superscripts, line breaks, and corresponding author information now follow MDPI Bioengineering style guidelines.

Revision:

  • Corrected author list, numbered affiliations, and correspondence formatting on the title page.

Comment 2: Abstract overstates readiness and impact

The abstract asserts broad transformative impacts without sufficient attention to validation and regulatory constraints.

Response:
We revised the Abstract to explicitly acknowledge validation, generalizability, governance, and regulatory challenges. Language implying immediate clinical readiness was replaced with a system-level framing that emphasizes current limitations and future requirements.

Revision:

Revised Abstract adding explicit references to validation gaps, equity concerns, and governance needs.

Comment 3: Hypertension call-out overstates cuffless BP maturity

Cuffless BP methods require calibration and lack robust validation.

Response:
We have removed this callout.

Revision:

Callout removed.

Comment 4: Inpatient / surgical AI conflates assistance with autonomy

Regulatory and safety distinctions between assistive AI and autonomous systems are not clear.

Response:
We clarified that current inpatient, procedural, and surgical AI systems are assistive rather than autonomous. The revised text explicitly distinguishes decision support and guidance from autonomous clinical action and references regulatory and safety constraints.

Revision:

Revised Section 3 (Inpatient and Acute Care), explicitly distinguishing assistive AI from autonomous systems.

Comment 5: Ambient listening lacks governance discussion

Consent, privacy, and legal governance are not sufficiently addressed.

Response:
We expanded the architecture section to explicitly address consent, privacy, data minimization, security, and governance considerations for ambient and conversational AI. The revised text emphasizes human oversight, accountability, and regulatory compliance.

Revision:

Revised Section 7 (Architecture and Systems), adding governance and privacy safeguards for ambient AI.

Comment 6: Bias and equity lack actionable mitigation strategies

Bias risks are described, but mitigation strategies and metrics are missing.

Response:
We strengthened the equity discussion by adding practical mitigation approaches, including dataset auditing, subgroup performance monitoring, and alignment with equity metrics. 

Revision:

  • Expanded equity discussion across Sections 1, 5, and 7

Comment 7: Explainability discussion is too generic

Interpretable models vs post-hoc explainability methods should be distinguished.

Response:
We revised the explainability discussion to distinguish inherently interpretable models (e.g., regression, decision trees) from post-hoc explainability techniques (e.g., SHAP, LIME), noting the limitations and appropriate use cases of each.

Revision:

  • Revised explainability and trust discussion in Sections 1 and 7.

Comment 8: Validation is discussed but not specified

Lack of specificity on how AI should be validated in healthcare.

Response:
We added concrete validation expectations, including prospective evaluation, external validation, human-in-the-loop testing, and post-deployment monitoring. 

Revision:

  • Expanded validation discussion across multiple sections

Comment 9: Interoperability lacks concrete standards

Interoperability is noted without reference to specific standards.

Response:
We agree and have revised Section 7 (Architecture and Systems) to explicitly reference widely adopted interoperability standards that enable AI-first healthcare systems. The revised text now names HL7 FHIR for clinical data exchange, DICOM for imaging interoperability, and the OMOP Common Data Model for analytics and population health applications. These additions provide concrete guidance on how interoperability can be operationalized across care settings and data types.

Revision:

Revised Section 7 to include concrete interoperability considerations.

Comment 10: Data privacy and security are superficial

AI introduces new attack vectors (e.g., model inversion, poisoning).

Response:
We agree and have revised Section 7 (Architecture and Systems) to explicitly acknowledge AI-specific security risks. The revised text now names data poisoning, model inversion, and adversarial inputs as distinct attack vectors introduced by machine learning systems, and emphasizes the need for security-aware AI deployment, monitoring, and governance beyond traditional health IT protections.

Revision:

Expanded security discussion in Section 7. This revision clarifies that AI security extends beyond conventional data protection to include model-level vulnerabilities.

Comment 11: Reproducibility and sharing

No statement about sharing code, models, or data.

Response:
We added language in the Conclusion emphasizing transparency, reproducibility, and the importance of sharing methods, models, or rationale for restrictions, consistent with emerging funder and journal expectations.

Revision:

Added reproducibility considerations to the Conclusion.

Comment 12–13: Typos, figures, and formatting

Multiple small errors and figure concerns noted.

Response:
All typographical issues, hyphenation inconsistencies, and caption errors were corrected. Figures that were unclear or potentially sourced externally were removed, and tables are now explicitly referenced in the text.

Revision:

Corrected typos and formatting throughout

Removed figures; ensured tables are referenced in text

Reviewer 2 Report

Comments and Suggestions for Authors

The authors provide a brief report on the integration of artificial intelligence (AI) in health care.

The report initially presents current implementation of AI in different parts of health care through a literature review which is adequate and generally up to date.

In the second part general requirements on the architecture and systems for a AI first healthcare system are presented. The authors analyse certain aspects and considerations for AI implementation and recognise challenges that have to be addressed.

The report summarises the current state of AI implementation in health care and proposes general guidelines for its full integration. The presentation is clear and comprehensive covering most aspects of the subject. I think the report is written in a way that can attract more interest from the general public, especially considering the two application examples presented.

Regarding figures and tables there is no reference to figure 1 and to the three tables in text. The tables provide useful information. The authors might consider commenting in text on figure 1.

In conclusion I think it is an interesting attempt to summarise the challenges of AI integration in healthcare that can be easily read by a wide number of readers.

Author Response

Response to Reviewers

We thank both reviewers for their careful reading of the manuscript and their constructive, detailed feedback. We have substantially revised the manuscript to address all comments. Below, we respond point by point to Reviewer 1 and Reviewer 2, indicating how each concern was addressed in the revised version.

Reviewer 2

General Comment

The manuscript is clear, comprehensive, and up to date; figures and tables should be referenced in text.

Response:
We appreciate the reviewer’s positive assessment. In response to the specific comment, we revised the manuscript to explicitly reference all tables within the main text and clarified their role in supporting the narrative.

Revision:

  • Table 1 has been revised and Tables 2-3 have been removed. Table 1 is referenced in the Conclusion.
  • Clarified their function in summarizing evidence, governance, and implementation readiness

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

The authors have incorporated all my concerns. There are no further requirements form my side.

Reviewer 2 Report

Comments and Suggestions for Authors

The authors have successfully addressed the suggested issues. I recommend the publication of the article in its present form.