AI-Enhanced Extended Reality for Rehabilitation in Africa: A Perspective on Explainable Agents, Co-Creation, and Generative Worlds

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

Review Report
AI-Enhanced Extended Reality Rehabilitation in Africa: A Promising Solution for Overcoming the Treatment Gap

This manuscript proposes an AI-enhanced extended reality (XR) rehabilitation framework tailored for low- and middle-income countries (LMICs), with a particular focus on Africa. The topic is timely and addresses an important healthcare gap; however, several substantive issues must be addressed to strengthen the rigor and clarity of the work.

Recommendation: Major Revision

1. Revise the abstract to clearly articulate the study’s objectives, methodology, key contributions, and practical implications.
2. Include a dedicated subsection (e.g., “6.5 Known Limitations and Failure Modes”) that critically examines risks associated with each AI technology, supported by relevant examples from medical AI literature.
3. Provide a technical specifications table or subsection outlining concrete model choices, hardware requirements, and expected performance. Alternatively, clearly state that implementation details are beyond the scope of this work and propose a benchmark study for future research.
4. Several claims (e.g., generative AI reducing development burden, large language models enabling real-time motion correction, explainable AI detecting bias) lack sufficient evidence. These should be supported with appropriate citations, rephrased more cautiously, or explicitly presented as hypotheses requiring validation.
5. Add a comparative subsection evaluating the proposed framework against existing low-cost rehabilitation solutions, with justification of its advantages in terms of cost-effectiveness, scalability, and clinical outcomes.
6. Expand the ethical analysis by including concrete recommendations related to governance structures, accountability, liability, and equitable access.
7. Revisit Section 4.3 to soften unsupported claims through appropriate hedging or by providing supporting references.

Addressing the above concerns will substantially improve the manuscript’s clarity, credibility, and practical relevance. Upon satisfactory revision, the paper could be considered suitable for publication.

 

Author Response

This manuscript proposes an AI-enhanced extended reality (XR) rehabilitation framework tailored for low- and middle-income countries (LMICs), with a particular focus on Africa. The topic is timely and addresses an important healthcare gap; however, several substantive issues must be addressed to strengthen the rigor and clarity of the work.

First, we thank you for your thorough and constructive feedbacks, which has helped us substantially improve the manuscript. Below we provide a point-by-point response to each comment, followed by the revised manuscript with all changes indicated.

Comment 1: Revise the abstract to clearly articulate the study's objectives, methodology, key contributions, and practical implications.

Response: We have completely rewritten the abstract to clearly state that this is a Perspective paper (not original research), specify our methodology (conceptual synthesis and framework proposal), and highlight the four key contributions: (1) identification of specific AI capabilities to address LMIC rehabilitation barriers, (2) a three-layer architectural framework, (3) implementation considerations for African contexts, and (4) a research agenda.

Comment 2: Include a dedicated subsection (e.g., "6.5 Known Limitations and Failure Modes") that critically examines risks associated with each AI technology, supported by relevant examples from medical AI literature.

Response: We have added a new subsection (Critical considerations) that systematically examines risks for each of the four AI technologies in Section 3: generative AI (hallucinated or unsafe content generation), LLMs (factual errors, inappropriate responses), MAS (unpredictable emergent behaviors), and XAI (gaming the explanation system, false reassurance). We also significantly developed the Section 6 about the risks and presented the mitigation strategies in Table 5

Comment 3: Provide a technical specifications table or subsection outlining concrete model choices, hardware requirements, and expected performance. Alternatively, clearly state that implementation details are beyond the scope of this work and propose a benchmark study for future research.

Response: We have added Section 5.6 "Technical Specifications and Benchmarking Roadmap" that provides concrete model choices (e.g., Stable Diffusion XL for 3D asset generation, Llama 3-8B quantized for on-device LLM, GPT-4 for clinician dashboard), hardware tiers (low-end, mid-range, high-end), expected performance metrics, and a phased benchmarking study. We summarized the key details in Table 3

Comment 4: Several claims (e.g., generative AI reducing development burden, LLMs enabling real-time motion correction, XAI detecting bias) lack sufficient evidence. These should be supported with appropriate citations, rephrased more cautiously, or explicitly presented as hypotheses requiring validation.

Response: We have systematically revised the manuscript to add relevant references to support our claims.

Comment 5: Add a comparative subsection evaluating the proposed framework against existing low-cost rehabilitation solutions, with justification of its advantages in terms of cost-effectiveness, scalability, and clinical outcomes.

Response: We have added Section 7 "Comparison with Existing Solutions and Value Proposition" that systematically compares our proposed AI-XR framework against: (1) conventional in-person therapy, (2) paper-based home exercise programs, (3) mobile health (mHealth) apps, (4) non-AI XR (e.g., AdaptRehab VR baseline), (5) tele-rehabilitation with remote clinicians, and (6) lower-cost 2D gamification. The comparison includes cost-effectiveness, scalability, clinical evidence strength, and implementation complexity. We try to present all the relevant information and technical details in Table 6.

Comment 6: Expand the ethical analysis by including concrete recommendations related to governance structures, accountability, liability, and equitable access.

Response: We have substantially expanded Section 6 (now reorganized and expanded) to include:

• New subsection 6.2 "Governance Structures"– proposing a multi-stakeholder governance board with specific membership and responsibilities
• New subsection 6.3 "Accountability and Liability"– discussing the "accountability gap" and proposing a tiered liability framework (hardware, AI model, clinical oversight, patient)
• New subsection 6.5 "Algorithmic Bias and Fairness"– as already requested and discussed
• Expanded equitable access discussionin subsection 6.7 "Equitable Access and the Digital Divide"

Comment 7: Revisit Section 4.3 to soften unsupported claims through appropriate hedging or by providing supporting references.

Response: We have thoroughly revised the whole manuscript to adopt a more conditional and neutral tone

Reviewer 2 Report

Comments and Suggestions for Authors

This manuscript addresses a timely and important topic, discussing the applications of AI-enhanced Extended reality for orthopaedic rehabilitation of patients in Africa. The paper is well-structured. However, as a conceptual framework, it would benefit from a more extensive discussion of several areas before publication.

Please refer to the recommendations below.

Major comments

1. There is insufficient clarity in the paper's genre and its related contribution. The paper is somewhere between a perspective piece and a framework proposal, without fully committing to either. Please clarify the genre more explicitly in the introduction. If you discuss a future perspective, your position needs to be more clearly stated and defended. If it is a framework proposal, the framework requires a more formal specification. Currently, Figure 1 presents the framework's architecture, but the three layers are described only indirectly, without details on how their components interact or how the system would be evaluated.
2. Some references provided in the paper weakly support the idea and should be used without overreliance. For example, the conclusions in reference 14 (AdaptRehab VR System), even if notable, do not provide sufficient grounds for extrapolation to a broader study, as it was based on a very small sample of 10 participants. Other LMIC-focused XR rehabilitation work should be reviewed more systematically to contextualise the gap the proposed framework aims to fill. Some references are unrelated to the healthcare industry and should be removed or replaced with more relevant ones. For instance, reference 19 cites gamification in business.
3. Sections 4.1-3.4 look overly optimistic about. They do not include sufficient critical discussion on the usage of Generative AI, LLMs, MAS, and XAI in the clinical context. For example, Generative AI for real-time 3D asset creation on edge devices remains technically challenging, LLMs in African languages are still significantly limited in quality and coverage, and the accuracy of LLMs as virtual advisors is questionable.
4. There is insufficient discussion of the ethical use of AI.
5. Section 7, the future directions/perspectives, is highly general. It would be better to separate them into near-term and longer-term categories, and to be more precise about the first.

Minor comments

6. The phrase "Bridging the Gap" in the title is used very frequently in global health literature. Consider a more specific one.

7. The abstract mentions the PESTEL analysis, but this is too far from the main contribution. Please consider whether it should be mentioned there.

Author Response

This manuscript addresses a timely and important topic, discussing the applications of AI-enhanced Extended reality for orthopaedic rehabilitation of patients in Africa. The paper is well-structured. However, as a conceptual framework, it would benefit from a more extensive discussion of several areas before publication.

First, we thank you for your thorough and constructive feedbacks, which has helped us substantially improve the manuscript. Below we provide a point-by-point response to each comment, followed by the revised manuscript with all changes indicated.

Comment 1: Insufficient clarity in the paper's genre and its related contribution. The paper is somewhere between a perspective piece and a framework proposal, without fully committing to either.

Response: We have clarified the genre explicitly in the title, in the abstract and Introduction. We now state clearly: "This paper is a Perspective that proposes a conceptual framework... It does not present primary empirical data but rather synthesizes existing literature and outlines a vision." We have also substantially expanded the framework specification with a detailed description of component interactions and an evaluation framework in Section 8.

Comment 2: *Some references provided in the paper weakly support the idea... For example, reference 14 (AdaptRehab VR) was based on a very small sample of 10 participants. Other LMIC-focused XR rehabilitation work should be reviewed more systematically.

Response: You are right, unfortunately currently there is still only very few works supporting the integration of AI for rehabilitation in Africa, therefore we have:

• Acknowledged the small sample size limitation of AdaptRehab VR explicitly in the text (Section 2, revised)
• Added some works performed in other emerging countries focused XR rehabilitation work, including studies from India, Brazil and China
• Removed reference 19 (gamification in business) and replaced with a clinically relevant citation on generative AI in healthcare [now ref 20]
• Reviewed all self-citations and reduced reliance on them; added 15 new external references

Comment 3: *Sections 3.1-3.4 look overly optimistic. They do not include sufficient critical discussion on the usage of Generative AI, LLMs, MAS, and XAI in the clinical context.

Response: We have substantially revised Sections 3.1-3.4 (now Section 3 with four subsections) to include:

• A dedicated "Critical considerations" paragraph in each subsection
• Specific technical challenges (e.g., generative AI real-time 3D generation on edge devices remains unsolved; LLMs in African languages have limited quality and coverage)
• Clinical safety concerns (e.g., LLM hallucinations, inappropriate recommendations)
• Citations documenting these limitations [31,34,35,56]

Comment 4: Insufficient discussion of the ethical use of AI.

Response: We have now substantially expanded Section 6 with new subsections on governance, accountability/liability, equitable access, and a consolidated summary table of risks and mitigation strategies.

Comment 5: *Section 7 (future directions) is highly general. Separate into near-term and longer-term categories, and be more precise.

Response: We have completely rewritten Section 8 to separate:

• Near-term (1-2 years):Specific pilot studies, dataset creation, lightweight model optimization
• Medium-term (2-4 years):Randomized controlled trials, multimodal interaction, integration with national health systems
• Longer-term (3-5+ years):Federated learning, foundation models for African rehabilitation, policy frameworks

Comment 6 (minor): The phrase "Bridging the Gap" in the title is used very frequently. Consider a more specific one.

Response: We have revised the title to: "AI-Enhanced Extended Reality for Rehabilitation in Africa: A Perspective on Explainable Agents, Co-creation, and Generative Worlds" (removed "Bridging the Gap").

Comment 7 (minor): The abstract mentions the PESTEL analysis, but this is too far from the main contribution. Please consider whether it should be mentioned there.

Response: We have removed the mention of PESTEL analysis from the abstract. The figure remains in the main text as a summarizing element.

Reviewer 3 Report

Comments and Suggestions for Authors

Reviewer's comments on the article: “AI-Enhanced Extended Reality for Rehabilitation in Africa: Bridging the Gap through Explainable Agents, Co-creation and Generative Worlds”

 

The manuscript proposes a conceptual AI-enhanced Extended Reality (XR) framework designed to address the shortage of rehabilitation professionals in Sub-Saharan Africa. The authors suggest integrating four AI pillars—Generative AI, Large Language Models (LLMs), Multiagent Systems (MAS), and Explainable AI (XAI)—to create culturally adaptive, personalized, and trustworthy telerehabilitation tools. The work builds upon the authors' previous experience with the AdaptRehab VR system and outlines a research agenda for implementing these technologies in low- and middle-income countries (LMICs).

While the topic is highly relevant and addresses a significant humanitarian and healthcare gap, the manuscript in its current form is strictly conceptual and descriptive. It functions more as a "Perspective" or "Position Paper" rather than "Original Research." The primary concern is the lack of empirical evidence, quantitative metrics, or a novel technical contribution that goes beyond synthesizing existing AI capabilities. Furthermore, the narrative is heavily centered on the authors' own ongoing projects, which may limit the objectivity of the proposed framework.

Key remarks and recommendations:

1. The manuscript exhibits a significant lack of scientific novelty and quantitative data, as it does not present any new algorithms, specific datasets, or results from pilot studies. There are no performance metrics regarding AI model accuracy in the African context or task success rates, and the "Methodology" section describes a literature selection process rather than a rigorous research framework.

Recommendations to authors: If the manuscript is to be considered as "Original Research," the authors must include empirical evidence, such as preliminary quantitative results from a pilot usability study or a simulation of the AI component’s performance, to move beyond a purely conceptual narrative.

2. There is an over-reliance on self-citation and personal projects, with approximately 15% of the bibliography consisting of the authors' own recent work (e.g., Kenea et al., 2025; Bonnechère, 2024/2025). The entire concept of the "AI-future" is constructed almost exclusively upon the base of the authors' AdaptRehab VR system, creating an impression that the manuscript is primarily designed to promote a specific ongoing project rather than providing an objective analysis of the broader technological landscape.

Recommendations to authors: The authors should  reduce the focus on their own publications and incorporating independent studies from other research groups working in similar resource-constrained environments (e.g., Southeast Asia or Latin America).

3. The manuscript suffers from methodological vagueness, as the selection of the four AI pillars (Generative AI, LLMs, MAS, and XAI) is not sufficiently justified through a comparative analysis or technical requirements gathering. The framework remains at a high level of abstraction, lacking the data flow diagrams or specific architectural constraints necessary for scientific reproducibility.

Recommendations to authors: It is recommended to provide a formal justification for the chosen AI technologies and include detailed architectural schemas or data flow diagrams that illustrate how these components interact within the proposed three-layer framework.

4. There is a clear contradiction regarding technological feasibility versus contextual constraints, specifically between the use of resource-heavy models (LLMs and Generative AI) and the mentioned infrastructural barriers like unstable power and limited internet. While "model quantization" and "edge computing" are mentioned, the paper lacks technical specifications or evidence that such complex systems can realistically operate on standalone XR headsets in rural settings.

Recommendations to authors: The authors should provide a technical deep dive into the "Decision Support Layer," specifying the parameters for the lightweight models intended for edge deployment and explaining the trade-offs between model complexity and the hardware limitations of the targeted geographical context.

Comments for author File: Comments.pdf

Author Response

The manuscript proposes a conceptual AI-enhanced Extended Reality (XR) framework designed to address the shortage of rehabilitation professionals in Sub-Saharan Africa. The authors suggest integrating four AI pillars—Generative AI, Large Language Models (LLMs), Multiagent Systems (MAS), and Explainable AI (XAI)—to create culturally adaptive, personalized, and trustworthy telerehabilitation tools. The work builds upon the authors' previous experience with the AdaptRehab VR system and outlines a research agenda for implementing these technologies in low- and middle-income countries (LMICs).

While the topic is highly relevant and addresses a significant humanitarian and healthcare gap, the manuscript in its current form is strictly conceptual and descriptive. It functions more as a "Perspective" or "Position Paper" rather than "Original Research." The primary concern is the lack of empirical evidence, quantitative metrics, or a novel technical contribution that goes beyond synthesizing existing AI capabilities. Furthermore, the narrative is heavily centered on the authors' own ongoing projects, which may limit the objectivity of the proposed framework.

First, we thank the reviewer for his thorough and constructive feedbacks, which has helped us substantially improve the manuscript. Below we provide a point-by-point response to each comment, followed by the revised manuscript with all changes indicated.

Comment 1: Significant lack of scientific novelty and quantitative data. If the manuscript is to be considered as "Original Research," include empirical evidence.

Response: We would like to clarify that this is not an Original Research paper but a Perspective paper. The revised title, abstract, and introduction clearly state this. However we know add more technical details and results, in particular we have added:

• A new subsection 5.6 with technical specifications and benchmarking roadmap
• Section 8: a clear Research Agenda with priorities and time line.

We believe the paper's contribution is conceptual synthesis and agenda-setting for an underexplored area (AI-XR in African LMICs), which is appropriate for a Perspective.

Comment 2: Over-reliance on self-citation and personal projects. Approximately 15% of the bibliography consists of the authors' own recent work.*

Response: We have:

• Reduced self-citations 7 out of 69 citations (which is relatively normal for a Perspective paper).
• Added 15 new external references, including independent studies from other research groups working in Southeast Asia, Latin America, and other African countries
• Added a review of other LMIC-focused XR rehabilitation initiatives in the introduction
• We add a complete section comparing existing solutions with our new proposed approach (section 7) and summarized the key elements in Table 6.

Comment 3: Methodological vagueness. The selection of the four AI pillars is not sufficiently justified. Framework lacks data flow diagrams or specific architectural constraints.

Response: We have:

• Explained in the introduction of Section 4 how we choose these 4 techniques and add some critical limitations in Table 1 and the rationale of using the different approach for rehabilitation in Table 2.
• Specified technical constraints(latency <100ms, offline operation, edge-cloud hybrid) in section 5.6 (technical specifications and benchmarking roadmap) (see comment 4)

Comment 4: Contradiction between resource-heavy models and infrastructural barriers. Lack of technical specifications for lightweight models on edge devices.

Response: We have:

• Added Section 5.6 "Technical Specifications and Benchmarking Roadmap"with concrete model choices for edge deployment:
  • LLM: Llama 3-8B quantized to 4-bit (~4GB) or TinyLlama-1.1B (~1GB)
  • Generative AI: On-device diffusion models (distilled versions, <2GB) or template-based procedural generation
  • XAI: SHAP/LIME lightweight implementations
• Added a table of hardware tiers and expected performance
• Explicitly discussed the trade-offsbetween model complexity and hardware limitations (also in Table 6).

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

The authors have carefully addressed all the reviewers’ comments and suggestions. The manuscript has improved significantly and is now suitable for publication in its current form.

Author Response

Thank you again for your comments that help us to improve the manuscript.

Reviewer 2 Report

Comments and Suggestions for Authors

Thank you for addressing all comments.

Author Response

Thank you again for your comments that help us to improve the manuscript.

Reviewer 3 Report

Comments and Suggestions for Authors

The authors have demonstrated a commendable effort in addressing the previous concerns. The reclassification of the manuscript as a "Perspective" paper significantly clarifies its scientific intent, and the reduction of self-citations (now ~10%) has greatly improved the objectivity and balance of the study. The addition of Section 5.6, with specific technical parameters for model quantization (e.g., Llama 3-8B, TinyLlama) and edge-deployment constraints, provides the necessary technical depth that was previously missing.
Specific remark regarding Figure 1: While the textual content has been significantly strengthened, Figure 1 remains unchanged. In its current form, the diagram is still overly abstract and does not fully reflect the newly added technical specifications, such as the data flow between edge and cloud, or the interaction between the specific AI pillars mentioned in the text.
Recommendation to authors: It is highly recommended to update Figure 1 to include a more detailed architectural schema. Specifically, the diagram should illustrate the data flow pathways and the technical integration points of the proposed models (e.g., where quantization occurs, how MAS interacts with the decision support layer). A more technical visual representation is essential to match the improved quality of the manuscript and to ensure the framework's clarity for future implementation.

Author Response

Thank you again for taking the time to revise this new version. According to your recommendations we updates Figure 1 to be aligned with the content of the paper and added a new figure (Figure 2) to represent the data flow