Empowering Medical Equipment Sustainability in Low-Resource Settings: An AI-Powered Diagnostic and Support Platform for Biomedical Technicians

AI summary

Overview

Research area: Artificial intelligence for global health — specifically LLM- and retrieval-augmented-generation (RAG) tools for maintaining medical diagnostic equipment in low- and middle-income countries (LMICs).

Technical level: Intermediate. The paper assumes familiarity with large language models, retrieval-augmented generation, and vector databases, but explains the architecture and evaluation in accessible terms.

Scope: The paper reports a completed Phase 0 proof of concept of the "INGENZI Tech" platform on a single device — the Philips HDI 5000 ultrasound machine — and lays out a five-phase roadmap toward multilingual, offline-capable, multi-device support (MRI, CT, X-ray) for biomedical technicians and imaging technologists/radiographers.

Authors and affiliations: Bernes Lorier Atabonfack, Ahmed Tahiru Issah, and Timothy X Brown (Carnegie Mellon University Africa, Kigali, Rwanda); Mohammed Hardi Abdul Baaki and Clemence Ingabire (Medical Artificial Intelligence Laboratory, Lagos, Nigeria); Tolusope Olusuyi (University of Pennsylvania); Maruf Adewole and Udunna C. Anazodo (McGill University, Montréal, Canada). Posted as arXiv:2601.16967v1 [cs.AI], 23 Jan 2026.

What This Paper Is About

In LMICs, a large share of medical diagnostic equipment sits idle or broken because maintenance is unstructured, trained biomedical engineering technicians (BMETs) are scarce, and devices acquired through third-party vendors or donations often come without manuals or manufacturer support. The authors build and test an AI assistant that lets a technician type in an error code or a symptom and get step-by-step troubleshooting guidance drawn from the device's own service documentation. A single ultrasound machine serves as the test case, with a peer-to-peer technician forum planned to capture the rare problems that documentation does not cover.

Key Contributions

1. An integrated LMIC-first diagnostic platform design. INGENZI Tech combines a multilingual, LLM-powered chatbot; a RAG architecture with segmented vector stores (user manuals, service manuals, and error codes kept separate) to reduce hallucination; offline-first deployment for low-bandwidth clinics; and a peer-to-peer technician forum. The authors position this as distinct from predictive-maintenance systems and manufacturer-side service tools that assume continuous sensor streams and reliable connectivity.

2. A validated Phase 0 proof of concept on the Philips HDI 5000 ultrasound. The prototype achieved 100% precision in error code matching and 80% troubleshooting accuracy, evaluated using internal evaluation methods.

3. A comparative landscape of existing AI medical-maintenance solutions (Table 1). The paper contrasts INGENZI Tech against Bruviti, Hadleigh Health / Vestfrost EMS / Nexleaf Analytics, the Bosnia AI-Metrology Project, Circuitry.ai, Stellarix / Kodexo Labs / Toronto Digital, and GE Healthcare Edison AI — noting that these are tied to proprietary hardware, online-only contexts, or high-resource health systems.

4. A five-phase development and evaluation roadmap. Phase 0 (proof of concept, completed), Phase 1 (forum integration and feedback loop), Phase 2 (API and IoT connectivity), Phase 3 (model optimization and continuous learning), Phase 4 (pilot deployment and evaluation in LMICs starting with East Africa), and Phase 5 (multi-device expansion beginning with Siemens MRI, CT, and X-ray). The paper states the project is "being carried out in 5 phases" while labeling them Phase 0 through Phase 5.

Main Findings

• Error code interpretation: 100% precision. All 90 error codes extracted from the Philips HDI 5000 ultrasound service manuals, submitted as queries with their descriptions, were retrieved correctly from the embedded vector stores. The authors attribute this to high alignment between query semantics and stored content.

• Instructional query evaluation: 80% accuracy. Of 30 natural language queries derived from the device's user and service manuals — covering operational instructions, safety guidelines, and device handling procedures — 24 received accurate and complete responses. The remaining six were partially accurate or missing critical details, which the authors trace to document chunking, relevance scoring, or retrieval granularity.

• Sub-10-second average latency. The system maintained this latency, which the authors describe as suitable for near real-time use in clinical environments.

• Dataset composition. Phase 0 used 15 technical documents related to the Philips HDI 5000 ultrasound system, parsed and chunked with a semantic-aware splitting strategy, embedded with OpenAI's GPT-3.5 Turbo embedding model, and stored in Facebook AI Similarity Search (FAISS).

• The LMIC equipment burden motivating the work. Studies cited estimate that 40% to 70% of medical equipment in LMICs is non-functional or underutilized at any given time. A Tanzania case study found 30–50% of medical equipment in sub-Saharan Africa experiences downtime. In Uganda, research across nine tertiary hospitals and five research institutions found 34% of medical devices faulty and 85.6% lacking operational manuals. The WHO estimates up to 70% of donated equipment fails to function as intended, and 40% of devices donated by Medical Surplus Recovery Organizations (MSROs) are inoperable. Device lifespan is often reduced by as much as 80%.

• Related predictive-maintenance work requires infrastructure LMICs often lack. Shamayleh et al. deployed an IoT maintenance system across over 8,000 medical devices in Malaysia with a 25% cost reduction using SVM-based failure classification; Zamzam et al. reported 99.4% prediction accuracy using sensor fusion across over 13,000 medical assets; Mohamed et al. proposed a digital twin for MRI equipment reducing downtime by 20%; and CNN-LSTM hybrid models have shown 92–99% accuracy in remaining useful life estimation. The authors argue these require continuous data streams, sensor infrastructure, and robust connectivity absent in many LMIC settings.

Methodology in Plain English

The project uses a phased, mixed-methods design combining iterative system development with quantitative performance evaluation and qualitative user feedback analysis. Only Phase 0 has been completed and reported.

For Phase 0, the team built a retrieval-augmented generation pipeline: a technician's query is interpreted through prompt engineering, the top-k most relevant document chunks are pulled from the vector stores, and those chunks are appended to the LLM prompt so the answer is grounded in actual service documentation. The model was GPT-3.5 Turbo. Documents were split with a semantic-aware chunking strategy to preserve contextual meaning. The vector database was FAISS, chosen for fast retrieval with minimal memory overhead, and was deliberately segmented into three separate stores — user manuals, service manuals, and error codes — so queries could be routed to the right domain and hallucination reduced. The backend ran on a lightweight Flask server with a React frontend, and the assistant supported custom tools for error code lookup, log parsing, maintenance schedule generation via the Google Calendar API, and simulated self-testing. Tool selection was driven by prompt conditioning and intent classification. Evaluation used internal methods: 90 error codes as structured queries and 30 natural language questions as unstructured queries.

Phases 1 through 5 are described as plans rather than completed work.

Why This Matters

Impact on research. The paper argues that most AI maintenance research assumes sensor infrastructure and constant connectivity that LMIC clinics do not have, and that little prior digital health work has addressed the operational maintenance of physical medical infrastructure. It contributes a design pattern — segmented RAG over proprietary service documentation, paired with human-in-the-loop peer knowledge capture — as an alternative to predictive maintenance for low-resource settings. It also targets human-centered AI in technical domains suffering from knowledge bottlenecks.

Real-world applications:

• A technician facing an unfamiliar error code on an unsupported or donated ultrasound machine can get a grounded, documentation-backed answer instead of taking the device offline indefinitely.
• Rural and low-bandwidth clinics can run offline-first assistance where internet access is unreliable, and non-English-speaking technicians can work in their own language.
• Biomedical engineering training programs and schools can use the tool as a teaching and reference aid, with Phase 4 planning onboarding materials and local technician training.
• A peer forum lets technicians share field-tested fixes for undocumented problems, capturing institutional knowledge that otherwise leaves when a technician does.

Industry relevance. The paper explicitly positions INGENZI Tech against commercial offerings (Bruviti, Circuitry.ai, GE Healthcare Edison AI, and IoT predictive-maintenance vendors), noting that those are largely tied to original equipment manufacturer ecosystems, enterprise clients, or continuous sensor data. An open, brand-agnostic architecture aimed at third-party and donated equipment addresses a market segment those solutions do not serve, and the planned expansion to Siemens MRI, CT, and X-ray points toward OEM-adjacent partnerships.

Future Directions

• Phase 1: Forum integration and feedback loop. Build or integrate a forum engine, link forum accounts with diagnostic tool profiles, design a schema for usage and feedback tracking, and pipe forum feedback into a structured pipeline for LLM fine-tuning and retrieval augmentation.
• Phase 2: API and IoT connectivity. Evaluate and implement DICOM, HL7, and MQTT; define secure API contracts for device integration; build a simulation layer for log ingestion without live devices; and prototype fault-alert dashboards for remote monitoring.
• Phase 3: Model optimization and continuous learning. Add a feedback classifier to label suggestions correct or incorrect, use forum responses and real interactions for semi-supervised learning, monitor hallucination frequency, optimize vector store indexing, and retrain against gold-standard troubleshooting guides. The paper leaves open how evaluation will use established LLM benchmarks adapted to technical reasoning and document retrieval, and what exact-match, precision, and recall targets will apply.
• Phase 4 and Phase 5: Deployment and scale. Pilot in East African clinical environments with partner hospitals and biomedical schools, measuring time-to-diagnosis, technician satisfaction, and device uptime improvement, then scale to MRI, CT, and X-ray starting with Siemens, adding device-detection logic and standardized ingestion formats. The paper states that physical repair execution, device manufacturing, and regulatory approval are outside scope and are "currently being investigated."

Target Audience

• Biomedical engineering technicians, imaging technologists, and radiographers working with donated or third-party medical equipment in LMICs.
• Health technology management and global health researchers studying equipment non-functionality, maintenance systems, and digital health infrastructure.
• AI researchers and engineers working on retrieval-augmented generation, domain-specific LLM grounding, hallucination reduction, and offline or multilingual deployment.
• Clinical engineering program designers, biomedical training schools, and NGOs involved in medical device donation and MSRO operations.
• Industry readers at medical device manufacturers and aftermarket service providers assessing service gaps for third-party and donated equipment in low-resource markets.

Authors’ abstract

In low- and middle-income countries (LMICs), a significant proportion of medical diagnostic equipment remains underutilized or non-functional due to a lack of timely maintenance, limited access to technical expertise, and minimal support from manufacturers, particularly for devices acquired through third-party vendors or donations. This challenge contributes to increased equipment downtime, delayed diagnoses, and compromised patient care. This research explores the development and validation of an AI-powered support platform designed to assist biomedical technicians in diagnosing and repairing medical devices in real-time. The system integrates a large language model (LLM) with a user-friendly web interface, enabling imaging technologists/radiographers and biomedical technicians to input error codes or device symptoms and receive accurate, step-by-step troubleshooting guidance. The platform also includes a global peer-to-peer discussion forum to support knowledge exchange and provide additional context for rare or undocumented issues. A proof of concept was developed using the Philips HDI 5000 ultrasound machine, achieving 100% precision in error code interpretation and 80% accuracy in suggesting corrective actions. This study demonstrates the feasibility and potential of AI-driven systems to support medical device maintenance, with the aim of reducing equipment downtime to improve healthcare delivery in resource-constrained environments.