Pretrained Biomedical Foundation Models: Comparative Analysis and Practical Applications in Healthcare

The rapid proliferation of pretrained biomedical foundation models has transformed healthcare AI, yet researchers and clinicians face persistent challenges in selecting the most appropriate model for a given application, as architectures, pretraining strategies, and computational requirements vary substantially across systems. This review provides a structured comparative synthesis of representative pretrained biomedical foundation models to support informed model selection for clinical and research applications. Seven models spanning encoderonly, decoder-only, and multimodal architectures and three pretraining strategies, continued pretraining, from-scratch pretraining, and instruction tuning, are compared across biomedical NLP benchmarks covering named entity recognition, question answering, and clinical reasoning. A multi-criteria framework evaluates each model across architecture, strategy, vocabulary type, task suitability, scale, and open-weight accessibility. Pretraining strategy and vocabulary alignment appear to be more influential factors in downstream performance than parameter count alone: from-scratch domain pretraining tends to outperform continued pretraining in several reported benchmarks, and instruction tuning has been shown to enable expert-level performance on specific clinical reasoning benchmarks (e.g., USMLE). These findings suggest that model selection should be guided by corpus-task alignment rather than scale, and that openweight alternatives can recover a meaningful fraction of closedmodel performance at lower deployment cost. Practical selection guidelines, current limitations including bias, hallucination, and computational accessibility, and future research directions are discussed.