[Paper Review] A Comprehensive Survey on Evaluating Large Language Model Applications in the Medical Industry

Yining Huang, Keke Tang+2 more|arXiv (Cornell University)|Apr 24, 2024

Artificial Intelligence in Healthcare10 citations

TL;DR

A survey compiling how large language models (LLMs) are evaluated in healthcare across clinical, data processing, research, education, and public health use cases, with discussion of benchmarks, metrics, and ethical challenges.

ABSTRACT

Since the inception of the Transformer architecture in 2017, Large Language Models (LLMs) such as GPT and BERT have evolved significantly, impacting various industries with their advanced capabilities in language understanding and generation. These models have shown potential to transform the medical field, highlighting the necessity for specialized evaluation frameworks to ensure their effective and ethical deployment. This comprehensive survey delineates the extensive application and requisite evaluation of LLMs within healthcare, emphasizing the critical need for empirical validation to fully exploit their capabilities in enhancing healthcare outcomes. Our survey is structured to provide an in-depth analysis of LLM applications across clinical settings, medical text data processing, research, education, and public health awareness. We begin by exploring the roles of LLMs in various medical applications, detailing their evaluation based on performance in tasks such as clinical diagnosis, medical text data processing, information retrieval, data analysis, and educational content generation. The subsequent sections offer a comprehensive discussion on the evaluation methods and metrics employed, including models, evaluators, and comparative experiments. We further examine the benchmarks and datasets utilized in these evaluations, providing a categorized description of benchmarks for tasks like question answering, summarization, information extraction, bioinformatics, information retrieval and general comprehensive benchmarks. This structure ensures a thorough understanding of how LLMs are assessed for their effectiveness, accuracy, usability, and ethical alignment in the medical domain. ...

Motivation & Objective

• Define the scope and need for specialized evaluation of LLMs in healthcare.
• Categorize LLM applications in medicine into clinical, data processing, research, education, and public awareness.
• Summarize evaluation methodologies, benchmarks, and indicators used across medical domains.
• Highlight challenges, governance, and strategies to improve evaluation frameworks for safe deployment.

Proposed method

• Surveyed literature and studies on LLM evaluations in medical settings across multiple domains.
• Organized the discussion by application field (clinical, data processing, research, education, public awareness) and by evaluation methodology.
• Summarized benchmark types and metrics used to assess accuracy, bias, safety, and clinical alignment.
• Synthesized findings to outline ethical, legal, and practical considerations for deployment.
• Provided guidance for practitioners, researchers, and policymakers on responsible evaluation and use of LLMs in healthcare.

Experimental results

Research questions

• RQ1 What are the predominant medical application areas where LLMs are evaluated?
• RQ2 What benchmarks, metrics, and evaluation protocols are used to assess LLMs in healthcare?
• RQ3 What are the major ethical, legal, and practical challenges in evaluating LLMs for medical use?
• RQ4 How can evaluation frameworks be improved to ensure safe and effective deployment in clinical settings?

Key findings

• LLMs have been evaluated across diverse medical fields, including general clinical tasks, specialty departments (e.g., endocrinology, ophthalmology), and radiology, with varying accuracy and biases reported.
• GPT-4 and PaLM-family models show strong performance on medical QA benchmarks (e.g., Flan-PaLM achieving 67.6% on MedQA), but human evaluation reveals concerns about clinical alignment and potential harm.
• ChatGPT variants achieve high accuracy in many clinical tasks but exhibit biases related to race, gender, and cost implications in care decisions.
• Multimodal medical LLMs (Med-MLLM) can perform radiology-related tasks with competitive results using limited labeled data (1%), indicating data efficiency advantages.
• Radiology and emergency medicine studies show promise for decision support and triage, but risk of unsafe recommendations and variable diagnostic accuracy requires careful governance.