Paper presents Cardiology Interface Terminology curated with machine learning

According to the arXiv paper (arXiv:2606.08311), the authors present a method to design a Cardiology Interface Terminology (CIT) intended for highlighting information in cardiology electronic health records (EHRs). The paper describes a three-phase process: building an initial CIT from cardiology-related SNOMED components and mined EHR phrases, semi-automatically reviewing candidate phrases to form a training-data CIT (TCIT), and training a machine learning model on TCIT to extract further concepts that form the final CIT. The paper reports that the highlighted test set attains 74.21% coverage and a breadth of 1.68, and that a 20-note sample gives average completeness 98.2% and conciseness 84.2% (arXiv:2606.08311). A conference chapter archived at ResearchWithNJ evaluates an ML-assisted curation variant and reports optimal manual-review batch sizes of 6,000 (concatenation) and 3,000 (anchoring), producing a CIT ML2+ with 68.74% coverage and breadth 1.6, close to a fully manually curated CIT+ (coverage 70.21%) (ResearchWithNJ). An IEEE conference paper from prior work outlines the same CIT concept for enabling fast skimming of cardiology EHRs and motivates annotation as distinct from standard named-entity recognition (IEEE).

The arXiv submission documents that the initial CIT is assembled from cardiology sub-hierarchies of SNOMED CT, additional SNOMED concepts mined from a build set, and domain-specific components such as abbreviations and medications. Candidate fine-grained phrases are extracted using concatenation and anchoring heuristics from the build set, then semi-automatically reviewed to create TCIT, which is used as training data for the machine learning extractor (arXiv:2606.08311). The ResearchWithNJ chapter states the ML model is a Neural Network trained on subsets of the candidate phrases to minimize manual-review workload while preserving highlighting quality (ResearchWithNJ).

Interface terminologies aim to provide higher-granularity phrases for downstream tasks like highlighting and summarization, a different objective than mapping text to reference ontologies. The reported workflow-seed with SNOMED CT, mine phrases from in-domain notes, then bootstrap an ML model using a semi-validated seed set-aligns with common patterns in clinical NLP for creating task-specific lexicons while limiting expert annotation costs. The ResearchWithNJ results illustrate a practical tradeoff: large but partial manual review (a few thousand phrases) can supply sufficient supervision for a model to approach the performance of fully manual curation.

Industry context: For practitioners building EHR-facing tools (summarizers, dashboards, clinical decision support), an interface terminology that prioritizes phrase-level granularity can materially change highlighting recall and usability compared to direct reliance on SNOMED CT or off-the-shelf NER. The reported coverage and completeness metrics suggest the approach captures a substantial fraction of clinically salient phrases in cardiology notes, though exact performance will depend on institution-specific language and note structure.

Follow-up work to validate CIT variants across multiple hospital systems and note types, peer-reviewed release of the annotated training splits or TCIT, and reproducibility artifacts (code, model weights) would determine how transferable the approach is for production clinical-NLP pipelines. Observers should also track comparisons against contemporary supervised NER and large-language-model augmentation approaches for EHR annotation.