From One-Size-Fits-All to Equitable, Context-Aware AI: Rethinking Digital Health in LMICs
Author: Yimam Getneh (PhD)
Introduction
Artificial intelligence is rapidly reshaping the future of healthcare. From diagnostic assistance and predictive analytics to clinical decision support and disease surveillance, AI is increasingly portrayed as a transformative force in global health systems (Topol, 2019; WHO, 2021). Across conferences, policy briefs, and technology forums, AI is often presented as a universally scalable solution that can improve care delivery anywhere. But beneath this optimism lies a questionable assumption: that healthcare environments are similar enough for the same digital systems to work everywhere. They are not. And even where they can work, they often work for some while excluding or harming others. Most AI-driven health tools are developed, trained, and validated in high-income settings, where stable infrastructure, electronic health records, specialist availability, and standardized workflows are taken for granted (Rajkomar et al., 2018). These conditions rarely reflect the realities of low- and middle-income countries (LMICs), where frontline care operates under fundamentally different clinical, operational, and infrastructural constraints. In many LMICs, health systems face persistent challenges: fragmented data ecosystems, intermittent internet connectivity, workforce shortages, limited diagnostic capacity, and diverse linguistic and cultural contexts (World Bank, 2023). Clinical decisions are often made under uncertainty, with constrained resources and variable access to specialized expertise. Yet despite these differences, digital health technologies are frequently transferred into these environments with minimal contextual adaptation. The result is a growing disconnect between technical capability and practical usability-and, more concerningly, a risk of exacerbating existing health inequities. The real challenge for global health AI, therefore, is not simply improving algorithmic performance. It is designing systems that understand the environments where healthcare actually happens, respect the priorities of local populations, operate within ethical guardrails that prevent harm, and actively work to close-not widen-health equity gaps. As AI becomes more integrated into care delivery, the central question is no longer whether AI can support health systems, but whether it can do so equitably, responsibly, and contextually across diverse settings (WHO, 2021). Digital health cannot succeed globally if its intelligence is built locally for only a small fraction of the world-or if it serves only the already-privileged segments within LMICs.
Approach: Equity as a Design Imperative
Rethinking digital health in LMICs requires moving beyond the idea that a single AI architecture can be universally deployed across fundamentally different health systems. Healthcare delivery is deeply shaped by context-by epidemiology, infrastructure, workforce composition, governance structures, language, and patterns of access (Bates et al., 2020). AI systems that ignore these realities risk becoming technically impressive but operationally ineffective-and in some cases, actively harmful or inequitable.
Context-aware and equity-centered AI offers a different path.
Rather than treating context or equity as secondary implementation challenges, this approach recognizes that local realities and existing power structures must shape how health technologies are designed, trained, validated, deployed, and evaluated. Such systems are built not only around data, but around the operational environments, ethical norms, disease burdens, and distributional consequences in which decisions are made. In LMIC settings, this requires rethinking several embedded assumptions in current digital health models:
1. Design for infrastructural variability
Continuous internet access, cloud-dependent architectures, and high-compute systems are not consistently feasible. Offline-capable, mobile-first, and low-resource-optimized systems become essential-not optional. Without this, rural and remote populations are systematically excluded.
2. Support frontline workers, don’t replace them
Many LMIC health systems rely heavily on community health workers, general practitioners, nurses, and non-specialist clinicians operating under significant workload pressure (Perry et al., 2014). AI must prioritize explainability, workflow integration, and decision augmentation over opaque automation. Critically, AI should not divert scarce training resources toward technology literacy at the expense of clinical skills-an equity risk often overlooked.
3. Align with local disease priorities
The burden of disease in many LMICs remains dominated by infectious diseases, maternal and neonatal conditions, malnutrition, and neglected tropical diseases-areas often underrepresented in commercial AI training datasets (Adamson & Smith, 2018). Models trained predominantly on non-communicable diseases in high-income populations may not only fail to generalize but also actively misdiagnose or overlook prevalent local conditions. Context-aware AI must be built around locally relevant epidemiological profiles, not convenient data sources.
4. Embed ethics from the ground up
Ethical AI cannot be an afterthought or a compliance checklist. In LMIC contexts, where regulatory oversight may be developing and power asymmetries between global tech developers and local communities are pronounced, specific risks arise: algorithmic bias reinforcing existing disparities, informed consent challenges in low-literacy settings, data privacy violations where legal protections are weak, and the potential for resource diversion away from proven public health interventions (Floridi et al., 2018).
5. Prioritize local validation, not just technical benchmarks
External validation on a held-out dataset is not enough. Models must be validated in situ—on local populations, across local infrastructure conditions, and against locally relevant outcome measures. Equally important: validation must be disaggregated by subgroups (gender, socioeconomic status, geography, language, disability, age) to detect where the model fails for specific populations.
6. Center equity explicitly.
Equity is not an automatic byproduct of good technical design. It requires active, intentional effort. This means:
· Access equity: Can the intended users-including the poorest, least literate, and most remote-actually reach and use the AI system?
· Fairness across groups: Does the model perform equally well across different demographic, linguistic, and clinical subgroups? If not, who is left behind?
· Benefit distribution: Who gains from AI deployment? If benefits accrue disproportionately to urban, wealthy, or formally employed populations, the AI is increasing inequity regardless of its technical performance.
· Mitigating digital determinants of health: Access to electricity, mobile networks, digital literacy, and affordable data plans are not evenly distributed. AI systems that assume these as givens will systematically exclude the most vulnerable.
Responsible AI as a Non-Negotiable Foundation
A recurring lesson from earlier digital health failures is that technical functionality does not guarantee ethical safety. In LMICs, where health systems are often overstretched and patients have limited recourse, the harms of a poorly designed or improperly deployed AI system can be severe: misdiagnosis, delayed care, loss of patient trust, or diversion of scarce resources toward ineffective tools.
Responsible and equitable AI in this context requires:
· Participatory design: Involving local health workers, patients, and policymakers-including representatives from marginalized groups-in the design process, not as consultants but as co-creators.
· Data sovereignty and representation: Ensuring that local populations retain ownership and control over their health data, with clear governance frameworks that prevent extractive practices. Also ensuring that training data represent all subgroups, not just the easiest to reach.
· Transparency and accountability: Clear documentation of model limitations, intended use cases, and performance boundaries-disaggregated by subgroup-made accessible to non-technical stakeholders.
· Ongoing equity monitoring: Continuous post-deployment surveillance for performance drift, bias, or unanticipated harms, with prespecified equity thresholds and clear pathways for feedback, recall, and remediation.
· Pro-poor design: Explicitly designing for the most marginalized user first-what works for a rural, low-literacy, elderly patient with a basic mobile phone will likely work for everyone else. Reverse innovation, not top-down optimization.
Without these safeguards, AI risks becoming another form of algorithmic colonialism-solutions extracted from and imposed upon vulnerable populations without meaningful benefit or consent, and often serving the privileged while neglecting the poor.
Insight: Equity as the True Test of Innovation
The future of digital health in LMICs may ultimately depend less on importing advanced technologies and more on redefining where innovation itself originates-and whom it serves. Too often, LMICs are positioned as downstream recipients of digital health solutions designed elsewhere. This framing overlooks a critical reality: many resource-constrained health systems have developed highly adaptive, resilient, and efficient approaches to care because of the constraints they face. These environments are not just deployment markets for AI-they are potential laboratories for the next generation of globally relevant equitable health innovation. Resource limitations have historically driven some of the most impactful forms of frugal innovation in global health. Mobile health systems, decentralized care delivery, task-shifting models, and community-centered approaches frequently emerged from LMIC contexts long before gaining wider international attention (Bhatti et al., 2018). These innovations succeeded not despite constraints, but because they were forced to solve for equity and access. AI development can follow the same trajectory. As healthcare AI evolves, future systems are likely to move toward contextual clinical copilots, syndromic intelligence platforms, multilingual decision-support tools, and adaptive surveillance systems capable of operating under constrained conditions. Advances in edge computing, lightweight models, and distributed AI architectures may further enable systems more responsive to local realities-not dependent on centralized infrastructure.
The most important perspective to add:
Equity is not a technical problem. It is a power problem. They require governance, accountability, and a redistribution of decision-making power in global health AI. The most valuable AI for LMICs may not be the most sophisticated algorithm, but the most equitable one. In settings where health disparities are already stark, an AI that works only for urban, literate, wealthy patients is not a partial success-it is a failure of justice. Conversely, an AI that meaningfully reaches and benefits the poorest, most remote, and most marginalized populations passes the only test that ultimately matters. The broader lesson is clear: intelligence alone is insufficient in healthcare. Relevance matters. Adaptability matters. Ethics matter. Local validation matters. Disease priorities matter. And equity matters most of all-because without it, AI becomes another tool for the privileged, leaving the already-left-behind even further behind. The future of digital health will not be determined solely by the scale or sophistication of algorithms-but by their ability to understand the environments where healthcare actually happens, the values of the communities they serve, the diseases that actually shape local mortality, and the equitable distribution of benefits across all segments of society. In global health, intelligence without context is approximation. Intelligence without ethics is dangerous. And intelligence without equity is injustice.
Reference
· Adamson AS, Smith A. Machine learning and health care disparities in dermatology. JAMA Dermatology. 2018;154(11):1247–1248.
· Bates DW, Levine D, Syrowatka A, et al. The potential of artificial intelligence to improve patient safety: a scoping review. NPJ Digital Medicine. 2020;3(54).
· Bhatti Y, Prime M, Harris M, Wadge H, McQueen J, Patel N. The search for the frugal innovation phenomenon in health care. NPJ Digital Medicine. 2018;1(14).
· Floridi L, Cowls J, Beltrametti M, et al. AI4People—An ethical framework for a good AI society: opportunities, risks, principles, and recommendations. Minds and Machines. 2018;28:689–707.
· Perry HB, Zulliger R, Rogers MM. Community health workers in low-, middle-, and high-income countries. Annual Review of Public Health. 2014;35:399–421.
· Rajkomar A, Dean J, Kohane I. Machine learning in medicine. New England Journal of Medicine. 2019;380:1347–1358.
· Topol E. Deep Medicine: How Artificial Intelligence Can Make Healthcare Human Again. Basic Books; 2019.
· World Health Organization (WHO). Ethics and Governance of Artificial Intelligence for Health. Geneva: WHO; 2021.
· World Bank. Digital Development Overview 2023. Washington DC: World Bank; 2023.
