Advancing Knowledge in Evaluating the Clinical Impact of Large Language Models for Clinical Text Summarization: A...

Area of Science:

• Clinical Informatics
• Artificial Intelligence in Healthcare
• Medical Documentation

Background:

• Large language models (LLMs) offer potential to streamline clinical text summarization from electronic health records (EHRs), reducing physician documentation burden.
• However, inaccuracies in LLM-generated summaries can lead to misinformation, misrepresentation of clinical facts, and potential patient safety risks.

Purpose of the Study:

• To systematically review and analyze the evaluation methodologies of clinical impact for LLM-based summarization systems.
• To assess how studies have examined utility, failure modes, patient safety risks, and bias in LLM-generated clinical summaries.

Main Methods:

• A literature search was conducted in PubMed for studies published between June 2024 and September 2025.
• A total of 144 studies were retrieved, with 32 selected for in-depth analysis based on predefined criteria.
• Evaluations were categorized across four dimensions: utility, failure, patient safety risks, and bias.

Main Results:

• Failure analysis (50%) was most common, focusing on inaccuracies, omissions, and hallucinations with varied definitions and methods.
• Utility analysis (25%) assessed workflow efficiency, readability, and understandability.
• Bias analysis (22%) examined gender, race, and stigmatizing language; patient safety risk analysis was minimal (3%).

Conclusions:

• Current evaluation frameworks for LLM summarization are evolving beyond technical metrics but lack standardization.
• Heterogeneous methodologies hinder consistent assessment of clinical impact and safe deployment.
• Future research requires standardized error taxonomies, risk assessment frameworks, and bias detection methods for reliable clinical integration.