The Specialist’s Paradox: Generalist AI May Better Organize Medical Knowledge

This study investigates the ability of six pre-trained sentence transformers to organize medical knowledge by performing unsupervised clustering on 70 high-level Medical Subject Headings (MeSH) terms across seven medical specialties. We evaluated models from different pre-training paradigms: general-purpose, domain-adapted, and from-scratch domain-specific. The results reveal a clear performance hierarchy. A top tier of models, including the general-purpose MPNet and the domain-adapted BioBERT and RoBERTa, produced highly coherent, specialty-aligned clusters (Adjusted Rand Index > 0.80). Conversely, models pre-trained from scratch on specialized corpora, such as PubMedBERT and BioClinicalBERT, performed poorly (Adjusted Rand Index < 0.51), with BioClinicalBERT yielding a disorganized clustering. These findings challenge the assumption that domain-specific pre-training guarantees superior performance for all semantic tasks. We conclude that model architecture, alignment between the pre-training objective and the downstream task, and the nature of the training data are more critical determinants of success for creating semantically coherent embedding spaces for medical concepts.