Echocardiography Report Translation and Inference Based on Parameter-Efficient Fine-Tuning of LLaMA Models.

Area of Science:

• Artificial Intelligence in Medicine
• Natural Language Processing for Healthcare
• Medical Informatics

Background:

• Echocardiography reports contain complex terminology hindering patient comprehension.
• There is a need for accessible medical information for non-specialists and patients.
• Current systems lack efficient translation and summarization for specialized medical reports.

Purpose of the Study:

• To develop a resource-efficient large language model (LLM) system for translating and summarizing echocardiography reports into Traditional Chinese.
• To enhance health literacy by creating patient-centric summaries.
• To address accessibility gaps in medical diagnostic tools.

Main Methods:

• Fine-tuning LLaMA-3.2-1B and LLaMA-3.2-3B-Instruct models using Quantized Low-Rank Adapter (QLoRA) and Unsloth acceleration on a single GPU.
• Implementing a dual-stage inference architecture for technical translation and patient-friendly summarization.
• Utilizing AdamW 8-bit optimizer for model training.

Main Results:

• The LLaMA-3.2-3B-Instruct model with AdamW 8-bit optimizer demonstrated stable validation performance.
• The system achieved strong results in semantic coherence and structural consistency based on BLEU, ROUGE, METEOR, and Perplexity metrics.
• Qualitative analysis identified areas for improvement, such as terminology simplification and minimizing hallucinations, highlighting the need for Human-in-the-Loop verification.

Conclusions:

• Deploying advanced medical AI in resource-limited settings is feasible.
• The developed platform provides a scalable foundation for improving clinical decision support and health literacy.
• Automated medical text processing can enhance the accessibility of diagnostic information.