INTRODUCTION The global discourse on artificial intelligence (AI) in healthcare often emphasizes resource-intensive infrastructure: Advanced graphics processing units, enterprise-level electronic health records (EHRs), and seamless highspeed connectivity. Yet across the healthcare landscapes of low- and middle-income countries (LMICs) – particularly in rural, mountainous, or overstretched primary care settings of South Asia – this high-tech paradigm remains aspirational rather than attainable. For district hospitals or remote health posts in India and Nepal, the pressing challenge is not a lack of enthusiasm for clinical innovation, but a chronic shortage of human resources, specialized diagnostics, and even reliable electricity. If AI is to serve as a catalyst for health equity rather than a driver of digital exclusion, it must pivot toward frugal innovation. Lightweight, localized, and mobile-driven AI systems – designed to function in resource-constrained environments – can empower frontline health workers, ease diagnostic bottlenecks, and reimagine medical delivery in clinics where conventional infrastructure is scarce. BRIDGING THE SPECIALIST GAP VIA POINT-OF-CARE DIAGNOSTICS The most immediate utility of AI in resource-limited settings lies in task-shifting and diagnostic triage. In South Asia, the patient-to-specialist ratio is critically skewed; rural areas rarely have permanent access to radiologists, dermatologists, or ophthalmologists. In these contexts, AI functions not as a substitute for clinical judgment, but as an accessible diagnostic bridge. Tuberculosis and respiratory screening Automated chest X-ray interpretation tools can detect pulmonary tuberculosis or acute pneumonia within minutes. For clinics equipped with basic digital X-ray machines but lacking radiologists, AI offers instant triage, flagging abnormalities for urgent referral. Diabetic retinopathy With the rising prevalence of the “South Asian phenotype” of diabetes, complications such as diabetic retinopathy remain a leading cause of preventable blindness. Deep learning models embedded in handheld, non-mydriatic fundus cameras enable community nurses to screen patients during routine visits. Validated systems – tested across primary care sites in India and Thailand – achieve specialist-level sensitivity, reducing the need for tertiary eye center referrals. Maternal health Lightweight, AI-augmented point-of-care ultrasound devices can guide minimally trained midwives through basic obstetric scans, automatically estimating gestational age and flagging high-risk fetal presentations. THE BLUEPRINT FOR FRUGAL AI ARCHITECTURE For AI to succeed in tier-2 or tier-3 clinics, systems must be engineered to withstand systemic constraints: • Mobile-First delivery clinical decision-support tools should run on personal mobile devices or low cost tablets. With smartphones already ubiquitous among South Asian clinicians, these platforms can host applications for automated image capture, documentation, and risk stratification. • Offline-First, lightweight models dependence on continuous cloud connectivity is impractical in regions prone to load-shedding or patchy cellular service. Advances in model distillation and neural network compression now allow computational demands to be reduced by up to 85%. This enables robust diagnostic models to run locally on smartphones or small servers, syncing to the cloud only when connectivity is available. • Contextual Integration Beyond EHRs Unlike Western clinics that rely on deeply integrated EHR systems, LMICs require AI tools that function as standalone workflows or link with flexible, open-source platforms such as OpenMRS – or even basic smartphone messaging applications. NAVIGATING RISKS: BIAS, ETHICS, AND THE TRAINING GAP While the potential benefits of AI in healthcare are immense, uncritical deployment carries significant risks. A central concern is algorithmic bias. Most globally available medical AI systems are trained on datasets from Western or East Asian populations. Variations in disease presentation, skin pigmentation (particularly in dermatology models), and local epidemiology mean these tools may underperform when applied to South Asian cohorts. Rigorous local validation on regional datasets is therefore an ethical imperative before widespread clinical rollout. Equally pressing is the training gap. Surveys reveal that although many medical faculty and students are aware of internet-based AI applications, few receive structured instruction on clinical AI – its utility, limitations, and risks, such as “hallucinations.” Without formal education, clinicians may treat AI outputs as infallible truths rather than probabilistic decision supports. This misinterpretation risks diagnostic premature closure and medical errors. CONCLUSION The integration of AI into resource-constrained clinics is no longer a futuristic luxury; it is a pragmatic necessity to counter severe workforce shortages. By prioritizing lightweight, mobile, and offline-capable tools, South Asia can leapfrog outdated diagnostic pipelines. Yet this digital transition must be anchored in updated medical curricula, rigorous local validation, and clear ethical guardrails. If pursued through the lens of frugal innovation, AI holds the power to democratize access to high-quality healthcare – reshaping clinical realities from Pokhara to the most remote corners of the globe.
Arun Kumar (2026) studied this question.