Empowering Remote Healthcare with On-Premises Solar-Powered AI Units: Design and Implementation

Abstract

Rural healthcare systems face considerable obstacles such as unreliable electricity, limited internet access, and shortages of healthcare professionals, all of which impede timely medical documentation and diagnostics. This study aims to design and evaluate a solar-powered AI unit equipped with fine-tuned Large Language Models for remote clinics, enabling offline medical transcription, clinical note generation, and diagnostic support in regions with limited infrastructure. Employing a mixed-methods approach, the research combines qualitative user experience assessments with quantitative performance metrics. Four TinyLLaMA models with 1.1 billion parameters were fine-tuned to generate Subjective, Objective, Assessment, and Plan (SOAP) notes using a synthetic dataset comprising thousands of patient records and transcriptions. These models were deployed on a Raspberry Pi 5, powered by solar panels, batteries, and a Wi-Fi antenna. System performance was simulated using mockup data, with plans for validation through real-world deployment. The fine-tuned models achieved high transcription accuracy, rapid note generation, and substantial diagnostic precision on mockup data, with a balanced demographic distribution. Qualitative feedback emphasized usability while highlighting challenges such as setup costs and the need for digital literacy. The solar-powered design ensures reliable offline operation, consuming roughly 480Wh daily. These solar-powered AI units and fine-tuned models present a sustainable solution to enhance documentation and diagnostics in remote healthcare settings. Real-world trials are crucial to validate system performance, complemented by strategic investments in training, infrastructure, and ethical governance to support scalability. This work has resulted in two provisional patent applications, further advancing its potential for practical deployment.