Digital rehabilitation in Africa could advance with AI and extended reality

Artificial intelligence (AI) and extended reality (XR) could help expand rehabilitation services across Africa, where stroke survivors and people with disabilities often face severe shortages of therapists, infrastructure and follow-up care, according to a new perspective paper. The authors insist that AI-enhanced virtual and mixed-reality systems could provide personalized, culturally adapted therapy in local languages, but warn that the technology remains unproven in African clinical settings and must be developed with strong ethical safeguards.

The paper, titled AI-Enhanced Extended Reality for Rehabilitation in Africa: A Perspective on Explainable Agents, Co-Creation, and Generative Worlds and published in Applied Sciences, proposes a conceptual framework that combines generative AI, large language models, multiagent systems and explainable AI with extended reality rehabilitation to address barriers in low- and middle-income African healthcare systems.

Africa's rehabilitation gap is pushing demand for scalable digital care

The need for rehabilitation is growing worldwide, but the shortage is particularly severe in low- and middle-income countries, where stroke, neurological injury, non-communicable diseases and disability are rising faster than health systems can respond. The authors point to Africa as a region where the mismatch between need and available services is particularly sharp. In parts of Sub-Saharan Africa, therapist availability can be extremely low, leaving many patients without sustained rehabilitation after stroke or other disabling conditions.

Upper limb rehabilitation after stroke is often essential for regaining independence, performing daily activities and improving quality of life. Without regular therapy, patients can face long-term functional limits that affect families, employment and community participation. Yet conventional rehabilitation models depend heavily on trained professionals, clinical infrastructure and repeated in-person visits. These requirements are difficult to meet in rural areas, under-resourced hospitals and communities where travel costs and distance block regular care.

XR technologies have already shown promise in rehabilitation. The term includes virtual reality, augmented reality and mixed reality systems that blend physical and digital environments. In rehabilitation, these tools can provide repetitive, task-based exercises, real-time feedback and game-like experiences that encourage motivation and adherence. For stroke survivors, virtual environments can be designed to train reaching, grasping, fine motor control and other upper limb functions.

The paper highlights the potential of serious games and immersive virtual reality to bring rehabilitation into homes, clinics and community settings. In principle, a patient could practice therapeutic movements using an affordable headset or device while receiving feedback from the system. This could reduce the burden on therapists and allow more frequent practice than traditional clinic visits alone.

But existing XR rehabilitation systems have major weaknesses. Many are developed in high-income settings and are not designed for African languages, cultural practices, local objects or infrastructure constraints. A game that uses unfamiliar environments or foreign-language instructions may fail to engage patients or may not fit local expectations of care. Systems also tend to offer limited personalization, basic feedback and manual content adaptation, meaning clinicians still need to spend time customizing exercises.

The authors cite AdaptRehab VR, an immersive virtual rehabilitation system developed through participatory co-creation in Ethiopia, as an example of how digital rehabilitation can be adapted to local context. The system used the Afaan Oromoo language and culturally familiar objects, including coffee beans, while targeting upper limb functions through virtual games. The project showed that culturally adapted XR can be acceptable to patients and clinicians, but it also revealed the limits of current systems: a small number of games, limited language coverage, manual customization and simple feedback.

This is where AI enters the argument. The authors argue that advanced AI could transform XR rehabilitation from a fixed set of exercises into a more adaptive ecosystem. Instead of relying only on pre-built content and therapist-led customization, AI could generate new therapy environments, provide conversational guidance, create virtual social support and explain clinical recommendations in ways patients and clinicians can understand.

The proposal is not presented as a finished clinical solution - it is a roadmap. The authors repeatedly stress that the evidence for AI-enhanced XR rehabilitation in African low-resource settings is still limited. The framework is designed to guide future development and testing, not to claim immediate readiness for large-scale deployment.

Generative AI, language models and virtual agents could personalize therapy

The proposed framework rests on four AI pillars, each aimed at a specific barrier to rehabilitation access. These include:

Generative AI

One of the major costs of XR rehabilitation is content creation. Developers must build 3D objects, virtual environments, sound files, instructions and exercises. For Africa, that challenge is multiplied by cultural and linguistic diversity. A rehabilitation system that works in one country, language or community may not work elsewhere without major adaptation.

Generative AI could reduce this burden by creating culturally relevant virtual content more quickly. A therapist or developer could request a virtual market scene, household setting or familiar object-based exercise, and the system could generate assets that fit local realities. Instead of asking patients to interact with unfamiliar Western objects, therapy could involve objects from daily life, including food items, tools or household materials.

This matters because rehabilitation is not only physical repetition. Motivation and meaning influence whether patients keep practicing. A familiar task can feel more relevant than an abstract digital exercise. Cultural adaptation may also improve trust, especially where imported digital tools are viewed with skepticism.

Large language models

These systems could support natural communication between patients and virtual rehabilitation platforms. A patient could ask why an exercise matters, receive simple guidance, hear encouragement in a local language or get a reminder about posture and movement. For community health workers, language models could provide just-in-time support when a specialist is not available.

In theory, this could help address the therapist shortage by offering basic coaching outside the clinic. But the authors are careful about the risks. Large language models can produce incorrect or misleading information. In rehabilitation, wrong advice about movement could cause injury or delay recovery. For that reason, the paper argues that clinical instructions should be constrained by templates, safety rules and clinician review. Open-ended AI responses may be safer for encouragement or general conversation than for unsupervised medical guidance.

Multiagent systems

Rehabilitation can be lonely, especially when patients practice at home without peers or group therapy. Virtual agents could act as coaches, peers or family-like supporters inside an XR environment. They could demonstrate exercises, encourage participation, support turn-taking or create group-like experiences for patients who are geographically dispersed.

Such systems could be especially useful where travel to rehabilitation centers is difficult. A virtual peer performing the same exercise could make therapy less isolating. A virtual coach could help maintain motivation. A virtual family member or culturally familiar social role could make the experience feel more natural.

Still, the paper warns that multiagent systems remain experimental in healthcare. Multiple AI agents can behave unpredictably, and poorly designed social cues could discourage patients or produce psychological harm. These tools would require careful testing, human oversight and safeguards.

Explainable AI

Trust is a key issue in medical technology. Patients and clinicians may resist AI recommendations if they do not understand why a system changed exercise difficulty, flagged a movement problem or suggested a new therapy task. Explainable AI would make the reasoning visible.

For patients, this could mean simple progress explanations, such as showing that shoulder movement improved enough to advance to a harder level. For clinicians, it could mean dashboards showing range of motion, task completion, compensation patterns, confidence estimates and recommended therapy adjustments. The paper emphasizes that AI-generated recommendations should be treated as suggestions requiring clinical review, not automatic treatment decisions.

This focus on explainability is especially important in low-resource settings, where clinicians may have limited exposure to advanced AI tools and regulators may lack detailed frameworks for AI medical devices. Clear explanations can support oversight, accountability and safer adoption.

The authors combine these four pillars into a three-layer architecture. A content creation layer would generate or adapt virtual environments and rehabilitation tasks. An interaction layer would manage patient coaching, dialogue and virtual social support. A decision-support layer would analyze movement data, generate explanations and help clinicians monitor progress. The system would ideally work on edge devices where possible, with periodic cloud synchronization to reduce dependence on unreliable internet.

Ethical safeguards and local co-creation will decide whether the technology works

AI-enhanced XR can democratize rehabilitation only if it is designed around African realities from the start, the authors point out. This means co-creation with patients, clinicians, community health workers, local developers and health authorities whilst acknowledging infrastructure constraints, including unreliable electricity, uneven internet access, limited device availability and weak technical support.

The framework recommends offline functionality as a key priority. Many rural or low-resource settings cannot depend on continuous cloud access. Systems should run locally on headsets, tablets or mobile devices and synchronize only when connectivity is available. The authors also propose different hardware tiers, from low-end standalone headsets using lightweight AI to high-end cloud-connected systems in better-resourced facilities.

Power supply is another barrier. Devices must be energy-efficient and compatible with battery or solar charging where electricity is unstable. Without this, even well-designed XR systems could sit unused in clinics that cannot support them.

Local capacity building is also important. A successful AI-XR rehabilitation system cannot depend permanently on foreign developers or external researchers. Local rehabilitation professionals, software developers and community health workers need training to maintain the systems, interpret outputs, manage risks and adapt content. The authors call for university curricula, train-the-trainer programs and open educational resources that build African expertise in AI and XR rehabilitation.

Cultural and linguistic adaptation must go beyond translation. AI systems should reflect local beliefs about disability, family involvement, communication styles and patient-clinician relationships. In some communities, family plays a central role in care. In others, gender, age or social hierarchy may affect how patients respond to virtual agents. A system that ignores these realities may fail even if the technology works.

Data governance is another major concern. Rehabilitation systems can collect sensitive health information, including movement patterns, functional limitations and patient behavior. In countries where data protection laws are still developing, technical safeguards become even more important. The paper recommends local data storage where possible, encryption, anonymization, strict access controls and clear patient consent.

The paper also identifies bias as a risk. AI models trained mainly on high-income populations may not perform accurately for African users because of differences in language, body types, movement patterns, health conditions and clinical settings. If systems are trained only on urban or educated users, they may perform worse for rural patients or people with low literacy. The authors call for diverse local datasets and regular bias audits.

Additionally, over-reliance on AI remains a key concern. In settings where rehabilitation professionals are scarce, governments could be tempted to use digital tools as substitutes rather than complements. The authors reject that approach. AI should expand access and support clinicians, not justify reduced investment in human rehabilitation workers.

To manage this risk, the paper proposes a human-in-the-loop model. Treatment recommendations should require clinician review, patients should be able to opt out of AI coaching, and high-uncertainty cases should be escalated for human attention. It also recommends governance boards involving clinicians, patients, family representatives, data protection officers, AI ethics experts, community health workers and government officials.

The paper outlines a phased research agenda. In the near term, researchers should create African rehabilitation datasets, fine-tune language models, test minimal AI-XR systems and validate safety. In the medium term, randomized controlled trials should compare AI-XR with conventional care and non-AI XR. These trials should measure functional recovery, adherence, cost-effectiveness, user trust and safety. In the longer term, researchers should explore federated learning, national scale-up models and regulatory guidelines for AI medical devices in low- and middle-income countries.

The authors also compare AI-XR with existing rehabilitation options. Conventional therapist-led care has strong evidence but low scalability in therapist-scarce regions. Paper-based home exercises are cheap but offer little feedback or personalization. Mobile health apps are more scalable but less immersive. Non-AI XR can improve engagement but still depends on manually developed content and therapist customization. AI-XR would be more complex and expensive at first, but could become valuable if it improves adherence, personalization and social support at scale.