The Temporal Validity of Taxonomic Entailments in Evolving Biomedical Ontologies

Abstract Background Biomedical knowledge-based systems, such as annotation pipelines, phenotype-driven diagnostic tools and cohort-selection and data-integration services, materialise conclusions derived from reference ontologies and act on them, yet treat the ontology as fixed although it is re-released every few weeks or months. How long such derived knowledge remains valid, and which maintenance strategies preserve its integrity, has rarely been quantified outside Gene Ontology enrichment analysis. This article quantifies both for the taxonomic fragment, treating each derived taxonomic entailment as a subject that is born when it is first derivable and dies when it no longer follows. Results From 311 release snapshots of eight biomedical reference ontologies of the Open Biological and Biomedical Ontology family spanning up to 16.3 years, 929,486 entailment lifetimes (623,167 incident) are reconstructed and analysed with Kaplan–Meier and stratified Cox models. Pooled five-year survival is 0.786 (95% ontology-level interval 0.708–0.886), or 0.608 counting loss of the justifying derivation; attrition is episodic and ontology-specific. Of the 128,950 hard invalidations, 71.8% arise from path breakage in which both classes named in the conclusion remain active. Deprecation signals on the two endpoint classes detect 14.5% of invalidations, and 16.5% when every intermediate class is watched as well; watching the subclass’s own parents raises this to at least 43.8%, whereas re-checking the edges of a stored justification path detects all of them. At workload level, 21.9% of 3,987 cohort-style subsumption queries lose at least one member of their answer set within three years. Volatility is largely unpredictable at derivation time: within-ontology models beat a permutation null in one of six evaluable ontologies, and cross-ontology transfer is below chance in five of seven. Conclusions Conclusions derived from biomedical ontologies expire at a measurable rate, and most of the failures are invisible to obsoletion-based monitoring. For taxonomic inference, consumers of these ontologies should re-derive conclusions on every release or store and re-check justification paths, rather than rely on deprecation signals, and ontology developers can help by publishing the entailments added and removed with each release. The corpus, the analysis code and the verification scripts are openly available.