Accurate discharge summary generation using fine tuned large language models with self evaluation.

Area of Science:

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

Background:

• Discharge summaries are vital for patient care but are time-consuming to create.
• Manual data integration leads to delays, errors, and omissions in documentation.

Purpose of the Study:

• To develop an automated framework for generating discharge summaries.
• To reduce clinician workload and improve documentation accuracy and completeness.

Main Methods:

• Utilized Decomposed Low-Rank Adaptation (DoRA) for fine-tuning large language models (LLMs) in the medical domain.
• Implemented a novel self-evaluation mechanism for iterative error detection and correction.
• Compared the framework against few-shot prompting and Chain of Thought (CoT) methods.

Main Results:

• DoRA outperformed traditional fine-tuning methods, improving BERTScore and reducing Perplexity.
• The self-evaluation mechanism significantly enhanced BERTScore (6.9% vs. few-shot, 4.1% vs. CoT) and ROUGE-L (69.6% vs. few-shot, 0.4% vs. CoT).
• Qualitative assessments confirmed improved accuracy and completeness of generated summaries.

Conclusions:

• The novel AI framework substantially improves discharge summary quality and consistency.
• Automated generation reduces creation time and enhances the reliability of clinical documentation.
• This approach shows significant potential for AI-driven healthcare documentation.