A dual-mode LLM framework for medical and general language translation for breaking barriers in healthcare communication

Abstract

Effective communication between healthcare providers and patients is frequently hindered by the complexity of medical language, contributing to misinterpretations and reduced health literacy. This study presents a robust, dual-mode translation framework powered by large language models (LLMs), combining neural architectures with rule-based safety layers to enable bidirectional translation between medical and general language. In addition to simplifying clinical language for patients, our system accurately reconstructs layperson descriptions into medically precise terms - empowering providers with clearer symptom narratives. We significantly enhance prior work by introducing domain-specific evaluation metrics (e.g., drug name preservation, dosage accuracy, ambiguity detection), conducting human-in-the-loop clinical validation, and benchmarking against specialized medical models such as MedPaLM, BioBERT, ClinicalBERT, and GPT-4 with medical prompting. Our curated dataset exceeds 60,000 annotated pairs across 15 specialties, ensuring generalizability across healthcare contexts. Clinical trials in outpatient settings demonstrate a 35% improvement in patient comprehension and high physician satisfaction (8.7/10), with zero safety incidents recorded. The system architecture supports real-world deployment via FHIR-compatible APIs and complies with regulatory frameworks such as HIPAA and FDA 510(k). These results indicate the model's potential to meaningfully bridge the communication gap in healthcare while setting a new standard for safe, transparent, and clinically grounded AI translation tools.