BioMed-LLM

Problem

General-purpose large language models struggle with biomedical language: they hallucinate clinical facts, misinterpret domain-specific terminology, and lack the structured reasoning needed for clinical decision support. Adapting foundation models for medicine requires both domain-specific fine-tuning and rigorous evaluation.

Approach

BioMed-LLM explores fine-tuning and retrieval-augmented generation (RAG) strategies for open-source LLMs on biomedical corpora. The project benchmarks several base models (LLaMA, Mistral, BioGPT) on clinical NLP tasks, then develops a multimodal extension that fuses radiology report text with imaging features from CIMILab's segmentation models.

Research directions:

• Instruction fine-tuning on curated biomedical QA datasets (MedQA, PubMedQA, BioASQ)
• RAG pipeline grounded in clinical guidelines and literature
• Multimodal fusion: text + imaging for joint report generation
• Hallucination detection and factual consistency evaluation

Results

• Fine-tuned models outperform zero-shot GPT-3.5 on PubMedQA by 7.3%
• RAG pipeline reduces factual hallucination rate by 42% vs. vanilla fine-tuning
• Multimodal extension (in progress) shows promising alignment between generated reports and ground-truth segmentations

Reproducibility

Datasets, fine-tuning scripts, evaluation harnesses, and model cards available in the repository following the CIMILab Reproducibility Standard.