Optimizing Edge AI for Rural Healthcare: A Case Study in Kenya

Thesis Abstract

The rapid expansion of telemedicine and digital health initiatives in Kenya’s rural regions has highlighted a persistent gap between centralized cloud-based AI services and the latency, bandwidth, and reliability constraints of remote health facilities. This study addresses the problem of delivering timely, accurate medical decisions in rural clinics by optimizing edge AI deployment to support diagnostic triage, remote monitoring, and decision support for frontline health workers. The aim is to design, implement, and evaluate an edge AI framework that maintains high diagnostic accuracy while meeting real-time constraints and energy efficiency in resource-limited settings. Specific objectives are (1) to identify clinical use-cases in rural Kenyan clinics where edge AI can reduce turnaround times and improve patient outcomes; (2) to develop an edge inference pipeline optimized for low-power hardware with constrained connectivity; (3) to evaluate model performance, latency, energy consumption, and robustness under heterogeneous network conditions; (4) to assess healthcare worker acceptance and workflow integration; and (5) to formulate a scalable deployment blueprint aligned with national digital health policies. The study adopts a pragmatic mixed-methods approach, combining quantitative performance evaluation with qualitative assessments. A multi-site study will be conducted in 12 rural health facilities across Kakamega, Kisii, and Turkana counties, serving an estimated population of 1.6 million people. The population comprises frontline clinicians, community health workers, and patients presenting with acute respiratory infections, maternal health concerns, and chronic disease management needs. A purposive sample of 48 clinicians and 240 patients will be recruited, with a subset of 16 clinics equipped with an edge AI-enabled decision-support device and 8 clinics serving as controls. Data collection instruments include (i) a standardized diagnostic accuracy and latency measurement protocol using 1,200 anonymized patient cases, (ii) energy consumption profiling on Raspberry Pi 4 and NVIDIA Jetson Nano platforms, (iii) workflow observation checklists, (iv) structured surveys and semi-structured interviews for clinicians, and (v) patient satisfaction scales. Validity and reliability are ensured through pilot testing, triangulation, and calculation of Cronbach’s alpha for survey instruments. Quantitative analysis employs regression analysis to examine relationships between edge inference latency, model accuracy, and clinical outcomes, while a repeated-measures ANOVA assesses differences in performance across sites and device configurations. Time-to-treatment metrics are analyzed using survival analysis where applicable. Model performance will be evaluated with standard metrics including accuracy, sensitivity, specificity, F1-score, and AUROC, alongside edge-specific metrics such as end-to-end latency, frames-per-second for streaming inputs, and energy-per-inference. An ensemble of lightweight CNNs and transformer-based models tailored for low-resource hardware (e.g., quantized models, pruned architectures) will be trained on Ghanaian and Kenyan health datasets, with transfer learning from publicly available radiology and dermatology datasets to bootstrap performance. The theoretical framework integrates the Technology Acceptance Model (TAM) to interpret clinician adoption, and the Resource-Based View (RBV) to understand organizational capability constraints impacting deployment. The study will also draw on the Distributed AI and Edge Computing for Healthcare theory to justify architectural choices. Expected findings include (i) edge AI can achieve diagnostic accuracy within 2–5 percentage points of cloud-based baselines while reducing mean latency by 40–60% and energy usage by 25–35%; (ii) robust performance under intermittent connectivity through local inference and opportunistic cloud synchronization; (iii) improved clinician workflow efficiency and higher patient satisfaction in intervention clinics; (iv) identifying organizational and policy prerequisites for scaling across rural systems. The contribution to knowledge lies in an empirically validated blueprint for edge AI deployment in rural low-resource health ecosystems, integrating architectural design, performance optimization, and user-centered implementation pathways. The study will provide policy-relevant recommendations for Kenya’s Ministry of Health on standards for data governance, interoperability, and training programs to sustain edge-enabled healthcare delivery. The main conclusion anticipates that carefully engineered edge AI pipelines, coupled with clinician engagement and supportive policy environments, can deliver timely, accurate, and acceptable diagnostic support in rural Kenyan health facilities, thereby narrowing the rural-urban health divide and informing scalable implementation models for similar settings across sub-Saharan Africa. Recommendations include scalable hardware procurement strategies, continuous model updates with federated learning, governance frameworks for patient data, and a phased rollout with monitoring dashboards to ensure ongoing performance and equity.