Clinical Note Generation From Doctor-Patient Conversations Using Parameter-Efficient Fine-Tuning Large Language...

Area of Science:

• Artificial Intelligence in Healthcare
• Natural Language Processing for Clinical Documentation
• Machine Learning in Medical Summarization

Background:

• Clinical note documentation is time-intensive, impacting healthcare efficiency.
• Limited open-source datasets hinder accurate clinical summarization using natural language processing.
• Large language models (LLMs) offer a promising solution for generating clinical notes.

Purpose of the Study:

• Evaluate parameter-efficient fine-tuned decoder-only LLMs for clinical note generation.
• Assess medical accuracy, robustness, and feasibility of PEFT under resource constraints.
• Compare LLM performance against traditional models for clinical summarization.

Main Methods:

• Fine-tuned decoder-only LLMs (Mistral, Meditron, Llama) using parameter-efficient fine-tuning (PEFT) on the Medical Training Summarization Dialog dataset.
• Evaluated models using ROUGE and BERTScore for automatic assessment.
• Expert physician review assessed clinical accuracy, completeness, concision, relevance, and readability.

Main Results:

• Meditron-7B and Llama3-8B achieved state-of-the-art results among open-source PEFT models.
• Decoder-only LLMs, especially Llama variants, outperformed traditional models.
• Human evaluation confirmed clinical coherence and accuracy of Llama3-8B and Mistral-7B.

Conclusions:

• PEFT of decoder-only LLMs can transform clinical workflows by streamlining documentation.
• These models provide a scalable and resource-efficient alternative for medical documentation.
• Optimizing LLMs through techniques like higher quantization can enhance efficiency without compromising performance.