AI and Healthcare: Building the Foundations of Intelligent Health Systems

Dr. Tantawy describes AI Share as an initiative developed by Egypt's Applied Innovation Center to enable countries in the Global South to share and co-develop AI health solutions rather than duplicating efforts. The platform acts as a 'basket of solutions' where entities can contribute tools they have developed, allow others to use and fine-tune them, and also signal problems they need help solving. This collaborative model is designed to maximise limited resources across the region.

Dr. Tantawy explained that countries in the South are resource-constrained and often develop solutions for the same problems independently, making collaboration essential . He described the platform as a shared repository where solution owners retain control but make their tools available for others to use or adapt, and where countries can also request co-development of solutions they cannot afford commercially .

Dr. Tantawy emphasises that AI in healthcare is not a single, monolithic concept but rather a collection of very different applications, each with its own requirements and challenges. He lists diagnostics, clinical decision support, administrative support, structured data capture from doctors, and research as distinct domains. Treating them uniformly would be a mistake, as each must be approached differently.

Dr. Tantawy explicitly enumerated the different domains of AI in healthcare, noting that the presentations from startup companies during the session illustrated this diversity clearly . He stressed that these domains are very different and must be looked at very differently .

Dr. Tantawy outlines a range of practical challenges that arise when deploying AI health solutions, beginning with the technical issue of data bias and representativeness, then moving to compute power, operational fit within existing hospital structures, staff training, and the high cost of per-patient pricing models for Southern countries. He argues that the technical side is actually the easiest to solve, while operational and cost challenges are often underestimated.

Dr. Tantawy noted that solutions developed in one country may perform poorly in another due to non-representative data, but that fine-tuning with local data can raise accuracy to 95-99% . He also highlighted that compute power is needed to improve both accuracy and speed , that operational integration into long-established hospitals is frequently overlooked , and that per-patient costs of even one or two dollars can be prohibitive for many Southern countries .

Dr. Tantawy argues that when AI is embedded in a health system, there must be guarantees that it will remain available and cannot be removed or misused in ways that harm patients. He frames this as a matter of national and humanitarian resilience, distinct from ordinary technology procurement. He calls for regulation that treats health AI as a humanitarian good rather than a standard commercial product.

Dr. Tantawy warned that introducing an AI solution into a healthcare system creates a dependency, and that nobody should be able to 'pull it out and cause harm to people' . He suggested this issue falls within the scope of regulation being developed by the UN and others, and that health AI should be treated as humanitarian in nature rather than like technology in banking or other sectors .

Dr. Tantawy draws on direct experience to show that AI models developed elsewhere often underperform when applied to a different population, but that fine-tuning with local data can dramatically improve results. This demonstrates that locally representative datasets are not merely desirable but essential for effective and safe AI deployment in health.

Dr. Tantawy recounted that Egypt had tested solutions developed in other countries and found their accuracy, specificity, and sensitivity to be inadequate, but that after fine-tuning with local data, performance rose to 95, 96, or even 99% . He also noted the importance of representing different age groups and genders within the data .

Chitose Noguchi describes UNDP's role as a facilitator of South-South cooperation, focusing on enabling the exchange of knowledge, approaches, and solutions rather than simply transferring finished products. She highlights that countries in the South share common challenges, priorities, and data issues, making this exchange particularly valuable. UNDP has worked to connect senior policymakers and regulators across multiple countries on responsible AI governance.

Noguchi cited a specific initiative with GSMA and the Ministry of Communications and Information Technology (NCIT) that brought senior policymakers and regulators from Egypt, Libya, Sudan, Tanzania, and Jordan together to work on responsible AI governance . She also noted that UNDP's facilitation focuses on common priorities and finding solutions that work in local contexts .

Noguchi describes the AI Hub for Sustainable Development as a concrete mechanism through which UNDP facilitates South-South cooperation by supporting AI entrepreneurs across Africa. Although physically located in Italy, the hub is designed to serve African countries and their innovators. Egypt's membership on the advisory board reflects the importance of Southern voices in shaping the hub's direction.

Noguchi explained that the AI Hub for Sustainable Development is housed in Italy but set up and facilitated by UNDP, with a focus on helping African countries adopt AI or support entrepreneurs developing AI solutions across the continent . Egypt sits on the hub's advisory board .

Noguchi argues that for AI to be adopted responsibly in health, countries must have robust governance and policy frameworks that embed transparency, accountability, and ethical principles. She stresses that healthcare is a domain where the stakes are especially high because AI decisions directly affect people's lives. Without such frameworks, the risks of harm and inequality are significant.

Noguchi referenced UNDP's 2025 Human Development Report and a related report on AI and inequality, both of which highlight that without the right policies and governance institutions, AI risks widening inequalities . She emphasised that governance frameworks must be particularly rigorous in healthcare given its direct impact on lives .

Noguchi argues that institutional capacity is a prerequisite for responsible AI adoption, meaning that public bodies must be equipped not just with technology but with the expertise to evaluate, procure, and oversee AI systems. UNDP has been actively building this capacity in Egypt and the region through training programmes and institutional partnerships. This includes work with the Egyptian Responsible AI Centre and the National Telecommunications Regulatory Authority.

Noguchi described UNDP's work with the Ministry of Communications and Information Technology to strengthen governance and institutional capacity, including training policymakers . She also mentioned collaboration with the Egyptian Responsible AI Centre and NTRA (National Telecommunications Regulatory Authority) as part of this capacity-building effort .

Noguchi identifies data governance as a foundational pillar for responsible AI in health, arguing that AI is only as good as the data underpinning it. She stresses that interoperability and responsible data sharing — especially across borders — are essential for enabling the kind of cross-regional collaboration discussed throughout the session. Without these foundations, equitable and sustainable AI deployment cannot be achieved.

Noguchi stated that data governance, interoperability, and responsible data sharing are critical, particularly for cross-border collaboration and partnerships . She also referenced the points raised by Dr. Amy Berk and Dr. Tantawy about data quality as a prerequisite for effective AI .

Noguchi raises the environmental dimension of AI deployment, noting that AI systems consume significant amounts of energy and that this is often overlooked in discussions about digital health. She argues that sustainability must be built into AI infrastructure from the outset, including through the development of green data centres. UNDP has been working with the Egyptian Ministry on this issue.

Noguchi noted that UNDP has been working with the Ministry on green data centres because AI uses a lot of energy, and that the connection between AI and environmental impact is not always made explicit but is very real .

Noguchi concludes with the overarching principle that AI must be human-centred and serve people's needs rather than becoming an end in itself. She warns that without appropriate policies and governance, AI has the potential to deepen existing inequalities between countries rather than reducing them. This concern is backed by UNDP's own research and reporting.

Noguchi referenced UNDP's 2025 Human Development Report, which focused on AI, and a related publication titled 'The Next Great Divergence: Why AI May Widen Inequality Between Countries', both of which highlight the risk of AI exacerbating inequalities in the absence of strong governance . She concluded that AI needs to serve people, not the other way around .

Darlington Akogo presents Moremi AI as a comprehensive, single-model AI system designed to handle a wide range of healthcare and biological tasks rather than being narrowly specialised. The system covers multiple imaging modalities, supports clinical decision-making, assists with patient triaging, and extends into life science research through an AI research assistant. This breadth distinguishes it from more narrowly focused diagnostic tools.

Akogo described Moremi AI as covering multiple radiology modalities including x-rays, CT, mammograms, and ultrasound, as well as oncology, maternal and neonatal care, dermatology, and orthopedics . He demonstrated its use for medical report generation across 30 imaging modalities , breast cancer detection via mammogram and ultrasound analysis , clinical decision support to reduce medical error , patient triaging using ESI level grading , and a life science research assistant called Moremi Co-Researcher capable of end-to-end genomics, transcriptomics, and proteomics analysis . The project received initial funding from the Gates Foundation .

Mina Shawky describes Clinidoo as a healthcare companion platform that solves the problem of patients not knowing where to seek care, particularly in underserved communities in Egypt and Africa. The platform uses a symptom-based algorithm to match patients with the right specialty and care modality, and supports voice input to make the service accessible to elderly users and those who cannot type. It has already served over one million patients.

Shawky explained that many patients in Egypt do not know where to go when they need care, leading to wrong diagnoses, unnecessary costs, and unresolved problems . Clinidoo addresses this by recommending the appropriate specialty and care type based on symptoms entered by the patient, including via voice input . The platform has served more than one million patients with the help of around 15,000 healthcare providers and 300 hospitals .

Shawky argues that the proliferation of AI-driven health tools risks becoming technology-led rather than problem-led, and that this is a fundamental mistake. He insists that the starting point must always be the real healthcare problem, not the technology. He also stresses that user experience is paramount — the technology should be invisible and seamlessly integrated into the patient's journey.

Shawky explicitly stated that 'AI is not the destination' and that 'solving the real healthcare problem is', describing this as the guiding approach behind Clinidoo . He also noted that for patients, what matters is the experience, and that technology should be invisible and smoothly integrated .

Dr. Reis argues that without deliberate South-South cooperation, the introduction of AI in health risks deepening existing inequalities, as lower- and middle-income countries have historically been disadvantaged in accessing new technologies. He identifies two specific equity risks: data colonialism, where Northern actors extract data from Southern countries without equitable benefit-sharing, and algorithmic bias, where tools trained on Northern data perform poorly or harmfully in Southern contexts. South-South initiatives are therefore not merely beneficial but necessary.

Dr. Reis drew on historical precedents such as HIV medicines and COVID vaccines to illustrate how lower- and middle-income countries are consistently disadvantaged in accessing new technologies, warning that AI could worsen health inequities if not managed carefully . He specifically named data colonialism - where Northern countries extract data without giving equal benefit back - and bias from using Northern data to develop tools for Southern populations as concrete risks .

Dr. Reis explains that WHO has been actively engaged in AI ethics and governance for health for over a decade, establishing a formal expert group in 2019 to develop ethical frameworks and guardrails. This work is driven by WHO's foundational mandate of health for all since 1948 and by the recognition that new technologies have historically bypassed lower- and middle-income countries. AI presents both a great promise and a significant risk of exacerbating these inequities.

Dr. Reis noted that WHO established an expert group on ethics and governance for AI in health in 2019 , and that WHO's concern for equity is rooted in its mandate since 1948 . He cited HIV medicines and COVID vaccines as historical examples of lower- and middle-income countries being disadvantaged in accessing new technologies .

Dr. Reis outlines a comprehensive set of risks associated with the use of large language models (LLMs) in healthcare, ranging from clinical risks such as hallucination and inaccurate outputs to systemic risks including de-skilling of health professionals, cybersecurity vulnerabilities, and the significant environmental cost of running these systems. He argues that these risks must be actively managed through governance frameworks.

Dr. Reis described the risk of hallucination - where AI systems generate false or made-up information - as particularly dangerous in health contexts involving serious diseases . He also identified the risk of de-skilling, where health professionals over-rely on LLMs and lose their own clinical judgement over time . He further cited cybersecurity risks from hospital systems being hacked , the large carbon and water footprint of LLMs , and concerns about compliance with existing data protection laws .

Dr. Reis argues that global ethical frameworks for AI in health are necessary but not sufficient; they must be translated into regional and national contexts through specific laws, regulations, and ethical guardrails. Without this localisation, global principles remain abstract and fail to protect people in diverse settings. WHO has been supporting this process through regional workshops.

Dr. Reis noted that WHO has conducted regional workshops on AI ethics and governance, including one in Cairo approximately three years ago involving all Arab countries and countries of the Emerald region . He stated that laws, regulations, and ethical frameworks need to be developed in every country, and that it is not enough to have global frameworks alone .

Dr. Shakweer frames the current moment as a critical inflection point where AI has moved beyond the experimental phase and is actively deployed within health systems, yet the governance, data systems, and institutional capacity needed to ensure it is safe and equitable are still under construction. This creates both a challenge and an opportunity, particularly for countries in the Global South to shape how these technologies are designed and governed.

Dr. Shakweer cited global evidence that AI is already being used across health systems for diagnostics, disease surveillance, and clinical decision support, with further expansion expected due to workforce shortages and rising demand . She also referenced the WHO's Global Digital Health Monitor, which shows that most countries are still at a mid-level of digital health maturity and that only a few have fully integrated systems .

Dr. Berk argues that the current inflection point in digital health is characterised by a fundamental shift from simply digitising information into electronic health records to actively using and optimising data with AI to generate meaningful clinical knowledge. This shift is enabled by emerging standards such as FHIR for data exchange and Clinical Quality Language (CQL) for deterministic clinical decision support. The result is a move from passive standardisation to semantic-driven intelligence.

Dr. Berk drew on her background as a nurse who witnessed early EHR adoption to contrast the previous era of passive digitisation with the current era of active, AI-driven data use . She specifically named FHIR as a data exchange standard and CQL (Clinical Quality Language) as enabling more deterministic and effective clinical decision support and quality improvement .

Dr. Berk argues that the quality and governance of data is the foundational prerequisite for effective AI in healthcare, and that AI cannot perform well if the underlying data is poorly understood, governed, or applied. She frames data understanding as the essential first step before AI can be layered on top. This principle applies across all AI use cases in health.

Dr. Berk stated explicitly that 'AI in itself is only as good as the data, because AI is grounded in the data', and that how data is used, understood, and applied is the precursor to execution . She also described the need to understand data semantically and layer AI on top of that understanding to make AI infrastructure optimal .

Dr. Berk intervenes to reinforce the point about de-skilling by drawing a parallel with the earlier adoption of electronic health records, where similar warnings were issued about over-reliance on technology. She argues that AI, like EHRs before it, is a tool and enabler, not a replacement for human clinical judgement. This lesson from the EHR era should inform how AI is introduced and governed in health systems.

Dr. Berk noted that during the EHR adoption era, there were explicit warnings not to let EHRs do the thinking for clinicians, and that the same principle applies to AI . She emphasised that AI does not replace human judgement or critical thinking .

Dr. Berk argues that the transition from AI pilots to full production deployment is a significant challenge that depends on three interconnected factors: trust among all stakeholders, shared incentives that make the AI meaningful and applicable to real problems, and shared standards that ensure consistency and scalability. She reiterates that AI must be problem-driven rather than technology-driven, echoing the point made by other panellists.

Dr. Berk described the work Microsoft is doing on digitising quality measures as a concrete use case grounded in shared standards, noting that quality measures impact everybody and that AI can serve as an enabler for scaling them . She emphasised that moving from pilot to production requires gathering trust, establishing shared incentives, and ensuring the AI is meaningful and purposeful . She also noted the importance of open-sourced technology and validation and benchmarking to ensure scalability .