A Narrative Review on the Application of Large Language Models to Support Cancer Care and Research

Ryzen Benson, Maurizio Elia, Benjamin Hyams, Ji Hyun Chang, Julian C. HongYearbook 2024 · DOI

TL;DR

Large Language Models (LLMs) like BERT and GPT are rapidly finding applications in cancer care, from aiding clinical decision-making and patient education to streamlining research processes. This review highlights their current promise and challenges, emphasizing the need for rigorous evaluation and the development of open-source models to foster wider adoption and trust.

What Problem Does This Paper Address?

Cancer is a complex disease with diverse patient needs and treatment strategies, making data analysis and utilization critical but challenging. While Natural Language Processing (NLP) has a history of assisting in cancer research and care, the advent of powerful LLMs like BERT and GPT has opened up new, more sophisticated avenues. This paper addresses the need to systematically review and understand the emerging applications of these LLMs in cancer care and research, identifying their current capabilities, limitations, and future potential. The authors aim to provide a narrative overview of how these models are being used to support clinical decision-making, educate patients and providers, and accelerate research activities.

Methods

This paper presents a narrative review of peer-reviewed articles published between January 2021 and December 2023, focusing on studies that applied BERT and GPT-based LLMs in cancer care, education, and research. The authors searched the Scopus database, excluding reviews and preprints, and screened titles and abstracts for relevance. Four authors iteratively categorized the selected articles into broad themes for qualitative analysis. They specifically excluded studies focused solely on bioinformatics, genetics, or genomics unless directly related to cancer care. The review emphasizes studies with well-described methods and results, acknowledging the exploratory nature of much of the current LLM research in this domain.

Datasets

Not explicitly detailed as distinct datasets for training or evaluation in the paper. The studies reviewed likely utilized various real-world clinical datasets, such as radiology reports, electronic health records, and patient-generated text (e.g., blogs, forum posts). The paper notes that many studies lack standardized evaluation metrics, and common computer science benchmarks like GLUE, MMLU, and SQUAD are not directly applicable to medical data due to privacy concerns and the inherent complexity of clinical care.

Key Insights and Ideas

  • LLMs are rapidly becoming versatile tools in oncology: From suggesting treatment plans to summarizing complex medical reports, LLMs are demonstrating a broad range of utility.
  • BERT and GPT architectures offer distinct advantages: BERT is favored for its fine-tuning capabilities in specialized tasks, while GPT models excel at generating coherent and human-like text for educational and informational purposes.
  • Clinical Decision Support (CDS) is a major application area: LLMs are being explored for treatment/screening recommendations, monitoring treatment response, and even emergency triaging.
  • Patient and provider education is a natural fit for LLMs: Their ability to answer questions, simplify complex information, and address myths is proving valuable.
  • Research acceleration is a significant benefit: LLMs are assisting with literature review, manuscript writing, cohort identification, and clinical data processing.
  • "Hallucinations" and lack of explainability remain key challenges: LLMs can generate incorrect or fabricated information, and understanding why they produce certain outputs is crucial for trust and safety.
  • Reproducibility and standardization are lacking: The non-deterministic nature of LLMs necessitates rigorous evaluation, and the absence of standardized metrics hinders direct comparison of studies.
  • Open-source LLMs are crucial for equitable access and transparency: While proprietary models like GPT are powerful, the development and adoption of open-source alternatives are vital for wider research and clinical use.

Key Results

The review identified 53 peer-reviewed articles, categorized into three primary themes:

  • Clinical Decision Support (39.6%): This included studies on treatment/screening planning (9 articles), treatment response/quality of life (5 articles), adverse event surveillance (5 articles), emergency triaging (1 article), and social support (1 article).
  • Cancer Education (28.3%): This encompassed patient & provider education (13 articles), addressing myths/misconceptions (1 article), and text summarization/simplification (1 article).
  • Research Support Activities (32.1%): This included idea development/manuscript writing (1 article), cohort identification/extraction (4 articles), information retrieval (7 articles), clinical data processing (2 articles), and NLP-centric tasks (3 articles).

Specific findings include:

  • GPT-4 generally outperformed GPT-3.5 in accuracy and minimized hallucinations in treatment recommendation tasks.
  • BERT-based models have shown promise in classifying metastatic disease response and extracting adverse drug events from clinical text, sometimes matching or exceeding human performance.
  • LLMs can answer radiation oncology questions with high proficiency, though they struggle with complex math-intensive queries.
  • While LLMs can answer patient questions accurately, their responses sometimes lack empathy and may not be easily comprehensible to non-experts.
  • Studies using LLMs for scientific writing found that while helpful for inexperienced researchers, they could also lead to acceptance of false information.

Limitations and Future Work

Limitations: * The review was limited to Scopus and excluded other databases like MEDLINE, potentially missing relevant computer science or NLP proceedings. * Only BERT and GPT-based LLMs were included, potentially overlooking other significant models. * The exclusion of reviews and preprint articles might have missed emerging trends or comprehensive overviews. * The search strategy was limited to "cancer" and "oncology," potentially excluding studies using more specific cancer terms.

Future Work: * Standardized Evaluation: Develop quantitative and qualitative metrics for evaluating LLMs in medical applications, considering privacy and clinical complexity. * Open-Source LLMs: Prioritize the development and deployment of open-source LLMs to enhance transparency, customization, and wider adoption in cancer research and care. * User-Centered Design: Incorporate patient, caregiver, and domain expert feedback in the development and evaluation of LLM-based tools, especially for patient-facing applications. * Human Oversight: Emphasize the critical need for human oversight of LLM-generated clinical information to ensure accuracy, completeness, and patient safety. * Regulation: Advocate for regulation of LLMs in healthcare to ensure data privacy, governance, and patient safety. * Reliability and Interpretability: Improve LLM reliability by training on clinically meaningful text and enhance interpretability through techniques like knowledge graphs and user-centered interfaces. * Integration: Explore the widespread integration of LLMs into EHR systems and clinical workflows.

Summary updated

Abstract

The application of LLMs in cancer care has shown promise across a variety of diverse use cases. Future research should utilize quantitative metrics, qualitative insights, and user insights in the development and evaluation of LLM-based cancer care tools. The development of open-source LLMs for use in cancer care research and activities should also be a priority.
BibTeX
@inproceedings{benson2024narrative,
  author = {Ryzen Benson and Maurizio Elia and Benjamin Hyams and Ji Hyun Chang and Julian C. Hong},
  title = {{A Narrative Review on the Application of Large Language Models to Support Cancer Care and Research}},
  booktitle = {Yearbook},
  year = {2024},
  doi = {10.1055/s-0044-1800726},
}
Figure