Role-Adapted Clinical Report Generation for Ultrasound Measurements in Low-Resource Settings

Abstract

Obstetric ultrasound is critical for monitoring fetal growth, yet in many low-resource settings, healthcare workers who perform or receive ultrasound measurements lack the training to interpret them clinically. We present a system that automatically generates role-adapted clinical reports from fetal biometry measurements, targeting six healthcare worker roles across three expertise levels. The system combines Retrieval-Augmented Generation (RAG) from a knowledge base extracted from the World Health Organization (WHO) Manual of Diagnostic Ultrasound with deterministic fetal growth percentile computation based on INTERGROWTH-21st international standards. The knowledge base is designed for multilingual extensibility: since the source material is from an official WHO document, entries can be translated into any target language by domain experts or machine translation services. A key design principle is that clinical decision support (red, yellow, and green alerts) is derived deterministically from percentile thresholds, not from the language model, ensuring safety regardless of LLM output quality. Evaluation demonstrates sub-millimeter accuracy in percentile computation, 100% correctness in decision support classification, measurable readability differentiation across roles (Flesch-Kincaid grade 8.8 for community health workers vs. 11-13 for clinical roles), and 98% factual consistency across 42 generated reports spanning seven clinical scenarios. The system is designed for local deployment without internet connectivity.

Resources

Details