Fine-Tuning, Retrieval-Augmented Generation, and Hybrid Adaptation of Language Models for Clinical Decision-Making in Health Care: Systematic Review.
Anshum Patel, Yugant Khand, Sai Krishna Vallamchetla et al.
✦ AI-curated · Sources linked
In 30 seconds
This systematic review evaluated the effectiveness of fine-tuning, retrieval-augmented generation (RAG), and hybrid adaptation strategies for transformer-based language models in clinical decision-making. Among 35 studies, fine-tuning achieved area under the receiver operating characteristic curve values up to 0.912 for cancer detection. The review highlights the need for alignment of adaptation strategies with specific clinical tasks.
Key findings
- Fine-tuning achieved area under the receiver operating characteristic curve values up to 0.912 for cancer detection.
- RAG improved guideline adherence from 71.1% to 92.1% in decision-support tasks.
- Hybrid systems exceeded 90% accuracy in external validation for complex clinical workflows.
- 25 studies were assessed as high risk for methodological limitations.
Why it matters
Understanding the comparative effectiveness of adaptation strategies for language models is crucial for their safe integration into clinical practice, ensuring that decision-support tools are reliable and aligned with clinical needs.
What to keep in mind
The review identified substantial methodological limitations, with 25 studies judged as high risk for bias and concerns regarding external validation and participant selection.