Exploring Large Language Models in Healthcare: Insights into Corpora Sources, Customization Strategies, and Evaluation Metrics

Shuqi Yang, Mingrui Jing, Shuai Wang, Jiaxin Kou, Manfei Shi, Weijie Xing, Yan Hu, Zheng ZhuarXiv 2025 · DOI · arXiv

Abstract: This scoping review synthesizes 61 recent studies (2021–2024) to map the landscape of Large Language Models (LLMs) in healthcare. It reveals a significant shift toward hybrid customization strategies—combining Retrieval-Augmented Generation (RAG) with fine-tuning—while highlighting a dangerous reliance on unverified data and a lack of standardized, clinically-grounded evaluation metrics.

Introduction: Bridging the "Trust Gap" in Clinical AI

While Large Language Models (LLMs) show immense promise for medical question-answering and decision support, they face a significant "trust gap" in clinical settings. The primary obstacle is model hallucination, with reports indicating that LLMs produce erroneous responses in up to 20–30% of certain clinical tasks.

Beyond accuracy, the field lacks a unified development roadmap. There is currently no consensus on which data sources (corpora) are most reliable, how to effectively customize base models for medicine, or how to measure success beyond general linguistic benchmarks. This paper systematically reviews the literature to identify best practices and critical gaps in the development of "Medical AI."

Methodology: A Systematic Scoping Review

The authors conducted a scoping review following the PRISMA-ScR (Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Reviews) guidelines. They screened over 14,700 records from major databases, including PubMed, Embase, Scopus, and Web of Science, to identify 61 high-quality studies published between 2021 and 2024.

The search strategy utilized a specific combination of free words and MeSH terms to capture the breadth of LLM research in the medical domain.

The PubMed search strategy utilized to identify relevant studies for the scoping review.

The review focused on three core pillars of LLM development: corpora sources, customization techniques, and evaluation metrics.

The Landscape of Medical Corpora

The data used across the 61 studies was categorized into four primary types:

  • Literature Materials (n=34): Textbooks, clinical guidelines, and peer-reviewed journals.
  • Open-source Datasets (n=33): Established benchmarks like MedQA-USMLE, PubMedQA, and MedMCQA.
  • Real-world Clinical Resources (n=24): Electronic Health Records (EHRs), diagnostic reports, and expert-drafted cases.
  • Web-crawled Data (n=11): Content from health forums and social media, which is often unverified.

A major finding was the geographic and socio-economic bias in training data. Most corpora are sourced from high-income regions (primarily the US and China), meaning these models may reflect biases or perform poorly when applied to diverse global populations or low-resource settings.

Customization Strategies: The Rise of the "Customization Cocktail"

The most successful models are no longer relying on simple fine-tuning. 41 out of 61 studies used a combination of techniques. The emerging "gold standard" involves using Parameter-Efficient Fine-Tuning (PEFT/LoRA) to teach the model specialized medical terminology, paired with Retrieval-Augmented Generation (RAG) to ground its answers in up-to-date, verified clinical guidelines.

The Evaluation Challenge: Beyond Linguistic Fluency

Many researchers still rely on general-domain NLP metrics like BLEU or ROUGE. While these measure linguistic similarity, they fail to capture "clinical semantic accuracy." A model could achieve a high ROUGE score while still recommending a lethal drug interaction. While 36 studies incorporated human-based evaluation, the authors noted that expert assessments can be subjective and biased without standardized frameworks.

Key Results and Model Dominance

  • Model Trends: The GPT series (n=26) and LLaMA series (n=15) are the most frequent base models, though domestic Chinese models like the Baichuan series are gaining ground.
  • Application Areas: Oncology and Ophthalmology were the most frequently studied medical domains.
  • Metric Trends: Accuracy/correctness is the most prevalent metric (n=20), followed by F1 score (n=15) and Precision/Recall (n=13).

Conclusion and Future Directions

To ensure the safe deployment of LLMs in healthcare, the authors propose a "reliability hierarchy" for training data: Level I (vetted guidelines), Level II (EHRs), and Level III (curated forum data), with dynamic weighting applied to each.

Future research must prioritize the development of standardized clinical benchmarks that emphasize safety, ethics, and differential diagnosis validity over mere conversational fluency. Moving beyond "in-silico" environments to real-world clinical validation remains the essential next step for medical AI integration.

Summary updated

BibTeX
@misc{yang2025exploring,
  author = {Shuqi Yang and Mingrui Jing and Shuai Wang and Jiaxin Kou and Manfei Shi and Weijie Xing and Yan Hu and Zheng Zhu},
  title = {{Exploring Large Language Models in Healthcare: Insights into Corpora Sources, Customization Strategies, and Evaluation Metrics}},
  year = {2025},
  eprint = {2502.11861},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2502.11861},
}
Figure