One-size-fits-all healthcare AI may deepen global health gaps

Artificial intelligence (AI) could help close major healthcare gaps in low- and middle-income countries, but only if tools are designed around the real conditions faced by clinicians in resource-constrained settings, states a new paper published in DIGITAL HEALTH. It warns that AI systems built for tertiary hospitals in high-income countries may fail, or even add risk, when deployed in clinics with workforce shortages, limited diagnostics, paper-based records and high patient volumes.

The article, titled "Designing AI tools to advance health equity in resource-constrained low- and middle-income countries," outlines eight design principles for clinician-facing AI tools in low-resource environments, focusing on problem-driven development, socio-technical context, appropriate task selection, point-of-care access, better decision-making, time savings, explainability and actionable recommendations.

AI must start with clinical problems, not available data

AI design in low-resource settings must begin with the health problem that needs to be solved, the paper stresses. In many settings, AI development starts where annotated datasets are available or where model performance can be optimized. Susanto and colleagues argue that this logic is often misaligned with the operational realities of resource-constrained healthcare.

In low- and middle-income countries, clinical settings face three major, interlinked challenges: limited access to care because of health worker shortages, weak infrastructure and the simultaneous burden of infectious and chronic diseases. These conditions make prevention and early detection especially important. AI tools should therefore be judged by whether they help address a defined care gap, not by whether they produce impressive technical scores.

The authors point to examples where AI tools have been applied to practical public health needs, including screening for infectious diseases such as tuberculosis and HIV, and chronic diseases such as diabetic retinopathy. In these cases, AI was evaluated in primary care settings to improve preventive care, increase productivity, reduce dependence on scarce specialists, improve timeliness and support cost efficiency.

This problem-driven approach also forces a comparison with non-AI alternatives. AI should be treated as one possible intervention among many. In some cases, better training, simpler clinical protocols, supply-chain improvements or workflow redesign may deliver greater benefit than an AI tool. When AI is appropriate, the model and interface must be selected according to the specific task and clinical environment.

The article also stresses the need to understand the socio-technical context before introducing AI. Healthcare systems are not just technical systems. They include people, tasks, environments, tools, workflows, regulations and organizational cultures. An AI model may perform well in isolation but fail when placed into a clinic where clinicians are overburdened, records are paper-based, internet access is unreliable or referral systems are limited.

AI should be designed and evaluated as part of a wider clinical work system. This includes studying how clinicians currently make decisions, what data they actually have, what equipment is available, how patients move through care and where delays occur. In resource-constrained settings, clinicians may rely heavily on experience or peer consultation because specialist tests and imaging are unavailable. AI must support this reality rather than assume conditions that do not exist.

AI should not simply automate whatever is technically possible, it should target tasks where support can improve decisions, reduce delays or expand access. Common AI tasks in healthcare include image interpretation, documentation and risk assessment. However, for low-resource settings, the right task depends on available data and local workflows. A tool requiring advanced inputs that clinics cannot collect will have little value.

Point-of-care tools must save time and improve decisions

The article points out that AI must be accessible where clinical decisions are made. In many low-resource settings, mobile devices are more available than desktop infrastructure or electronic medical records. Clinicians already use smartphones for communication, consultation and clinical decision-making. For that reason, mobile-based AI tools may be especially useful if they are simple, offline-capable and usable by clinicians with basic digital literacy.

Accessibility, however, is not enough. AI must improve decisions. The authors argue that an AI-assisted clinician should perform better than the same clinician working unaided. Comparing an AI model against clinicians alone is not sufficient because most contemporary AI tools are assistive. They do not replace the clinician. They shape the interaction between human judgment and machine recommendation.

The article calls for evaluations that compare AI-assisted decision-making with standard clinical practice. Randomized controlled trials can assess real-world effects, but clinical vignette studies can offer safer, earlier testing before costly or disruptive implementation. This matters because poorly tested AI in healthcare can introduce patient safety risks.

Time is another major factor. In low-resource settings, clinicians often face heavy patient loads and very limited consultation time. AI tools that require extra review, data entry or verification may slow care and become unusable. The article warns that assistive AI must not add workload or delay diagnosis and treatment. It must provide timely recommendations that make time-sensitive decisions faster, not more cumbersome.

The authors cite evidence that AI can reduce diagnostic time in some contexts, including childhood cataract screening, and highlight the potential of AI-assisted ECG triage to accelerate care for patients with myocardial infarction. But the broader point is that time cost should be measured, not assumed. An AI tool that saves time in one setting could create clumsy automation in another if it disrupts workflow or demands additional steps clinicians cannot absorb. This is particularly important where manual documentation remains common.

Without electronic medical records, clinicians may already spend time gathering and recording information manually. AI tools must therefore be designed around the minimum data required for meaningful decision support. They should also fit naturally into clinical routines, even if some workflow adjustment is necessary.

The article calls this mutual adaptation. AI tools may require more structured data entry or more consistent documentation, but workflows may also need redesign so the tool can be used safely. The goal is not to drop AI into existing practice unchanged. The goal is to integrate it into a socio-technical system where both the tool and the workflow support better care.

Explainable and actionable AI is essential for health equity

The article warns that AI recommendations must be clinically comprehensible. In high-resource settings, clinicians may have access to specialist opinions, imaging, laboratory tests and other diagnostics to verify AI outputs. In low-resource settings, those confirmatory options are often limited. That makes explainability more important.

Clinicians need to understand why an AI tool produces a recommendation. A black-box risk score is not enough. AI outputs should show relevant input factors, risk drivers and reasoning cues that help clinicians interpret results in the patient's context. This preserves clinical discretion and supports safe decision-making. It also helps clinicians explain decisions to patients, which is essential for shared decision-making and trust.

The authors note that clinicians are more likely to accept AI when outputs are understandable and clinically meaningful. Explainability should not be designed for engineers alone. It should support the way clinicians reason under real-world constraints.

Actionability is equally important. AI tools that predict risk without offering realistic next steps provide little value. A recommendation must lead to something clinicians can actually do in that setting. If a tool identifies a patient as high-risk but the clinic lacks medicines, referral pathways or equipment, the output may not improve care. It may even increase frustration and ethical burden.

The article uses cardiovascular disease risk assessment as a case example. In resource-constrained primary care settings in Indonesia, doctors play a key role in preventing atherosclerotic cardiovascular disease by assessing and managing risk factors. The authors describe work showing that an AI prototype for cardiovascular risk assessment improved risk classification, reduced decision-making time and increased statin prescribing and referrals. This example demonstrates how the eight principles can be applied to a defined clinical problem.

The case also shows why actionability must be grounded in local resources. Cardiovascular risk prediction is useful only if clinicians can respond with feasible interventions, such as prescribing statins or making appropriate referrals. AI implementation should therefore be guided by local clinical consensus, policies and ethical oversight. Governance is needed to ensure transparency, accountability and responsible use.

AI can advance equity only when it expands useful care for underserved populations rather than widening the gap between well-resourced and under-resourced systems. If AI tools require advanced infrastructure, specialist interpretation, constant connectivity or expensive proprietary platforms, they may reinforce inequity. If they are designed for point-of-care use, aligned with local workflows and focused on feasible action, they may strengthen frontline care.

The authors also acknowledge that their principles need further testing across different countries, clinical governance structures and AI applications. The cardiovascular case is based on Indonesia and focuses on one clinical task. More research is needed to determine how these principles generalize across other low- and middle-income settings, diseases and healthcare systems.