Standard coherence metrics for biomedical topic models encode no clinical knowledge and cannot detect clinically implausible topic groupings. We propose SNOMED CT Wu?Palmer hierarchy distance as a post hoc, ontology-grounded diagnostic. On vascular surgery (47,318 articles) and craniofacial surgery (27,493 articles) corpora, the metric flags clinically heterogeneous topics that coherence misses?e.g., abdominal aortic aneurysm repair grouped with deep vein thrombosis (d = 0.600). Diagnostic signals are nearly identical across eight BERTopic embedding strategies including ontology-enhanced models, but diverge across model families: BERTopic alone produces a positive within- vs. cross-topic Cohen’s d, while LDA, NMF, and Top2Vec at matched topic counts score below their own cross-topic baselines (Cohen’s d 0; Mann?Whitney p 0.99). The score is therefore sensitive to topic-model output choice, not only to embedding choice within a single pipeline. A pre-clustering screening experiment finds near-zero correlation (|?| 0.08) between embedding cosine and SNOMED CT similarity, arguing that ontological validation belongs after clustering rather than as an embedding screen. We additionally describe a two-stage UMLS-CUI stopword filter that preserves high-frequency domain-specific concepts which naive frequency filtering would discard. After one-time concept curation, the diagnostic itself is automated and requires no per-topic expert scoring.