Do Medical LLMs Need Medical Pretraining? Evidence from Breast Cancer Text Classification

Résumé fourni par la source

Breast cancer diagnosis requires accurate classification tools to support clinical decision making. This study evaluates four large medical language models (LLMs) for automated breast cancer classification using parameter-efficient finetuning. Qwen2.5-7B, OpenBioLLM-8B, BioMistral-7B, and Meditron-7B are assessed using Quantized Low-Rank Adaptation (QLoRA). All models are finetuned with consistent hyperparameters targeting attention and feed-forward layers. The results demonstrate that the instruction-tuned models outperform the domainspecialized base models. Qwen2.5-7B and OpenBioLLM-8B both achieved 98.46% accuracy with F1-macro scores of 98.75% and 98.30% respectively, while BioMistral-7B reached 89.23% accuracy. Meditron-7B, despite medical pretraining, achieved only 53.85% accuracy. These findings indicate that the ability to follow instruction and general reasoning skills is more critical than domain-specific pretraining for medical text classification tasks. The results suggest that resource-efficient LLMs can provide reliable breast cancer classification, allowing deployment in resource-constrained clinical settings while maintaining high diagnostic precision.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.