TL;DR
This paper proposes a novel method to semantically categorize clinical terms by analyzing their surface-level features within different sections of Electronic Health Records (EHRs). By leveraging "sublanguage" properties – the unique linguistic characteristics of specific document sections – the authors demonstrate that these features can effectively predict a term's semantic category, offering a way to enrich existing medical terminologies.
What Problem Does This Paper Address?
The automatic processing of clinical documents, like Electronic Health Records (EHRs), is crucial for advancing healthcare. A key challenge is linking the vast array of clinical terms encountered in practice to established medical concepts within terminologies such as SNOMED CT. Existing terminologies often lack the full spectrum of term variants used by clinicians, especially in languages other than English. While machine learning approaches are popular, they often rely on pre-compiled terminologies. To bridge this gap, researchers aim to enrich these terminologies by identifying and categorizing new term variants. This paper specifically tackles the problem of semantic categorization of these term variants, focusing on how the unique linguistic characteristics of different parts of clinical documents (sublanguages) can be used to achieve this.
Methods
The core of this research lies in using "sublanguage features" to predict the semantic category of clinical terms. Instead of relying on the meaning of the words themselves or their surrounding context, the authors analyze the formal properties of the terms' surface form. These properties are derived from the observation that different sections within EHRs exhibit distinct linguistic styles and functions (e.g., "Complaints" section is narrative and uses lay terms, while "Medication" is a list of drugs).
The method involves:
- Feature Engineering: Identifying and annotating terms with binary features that capture their formal properties. These include:
- REGISTER: Whether the term is a standard term (as in SNOMED CT), an abbreviation/acronym, or a derivation/paraphrase/compound.
- REDUCTION: Whether the term is a shortened form.
- MORPHO-SYNTACTIC VARIANT: Whether the term deviates from standard grammatical or syntactic structures.
- Sublanguage Analysis: Recognizing that different sections of EHRs (e.g., Anamnesis, Comments, Complaints, Conclusion, Examination, History, Medication, Therapy) act as distinct sublanguages, each with its own stylistic properties.
- Classification: Employing a Random Forest classifier to predict the semantic category of a term based on its formal features and the section of the EHR it appears in. The hypothesis is that the informativeness of these formal features will vary depending on the sublanguage characteristics of the section.
Datasets
The study utilizes a corpus of 4,426 Electronic Health Records (EHRs) written in Belgian Dutch, provided by a Belgian hospital. These EHRs pertain to patients diagnosed with diabetes undergoing routine check-ups.
- Corpus: The EHRs were de-identified and structured into distinct sections, each with a specific communicative function and style.
- Term Annotation: In a prior project, terms from this corpus were manually annotated with concept codes from SNOMED CT. After validation, 15,025 unique terms linked to 7,687 concepts remained.
- Semantic Grouping: Concepts were mapped to UMLS semantic groups.
- Formal Annotation: For this study, the unique terms were manually annotated with the formal features described in the Methods section.
- Concept Selection: For the classification task, five concepts were chosen from each of the five most frequent semantic groups (DISORDERS, PROCEDURES, CONCEPTS & IDEAS, CHEMICALS & DRUGS, ANATOMY). The selection aimed for diversity in formal alternations and coverage of the semantic class. Concepts occurring less than 500 times within a section were excluded.
Key Insights and Ideas
- Sublanguage Matters: The unique linguistic characteristics of different sections within clinical documents (sublanguages) significantly influence the effectiveness of formal features for semantic categorization.
- Surface Form is Informative: Even without understanding the meaning of a term or its context, its formal properties (like being an abbreviation or a compound word) can be strong indicators of its semantic category.
- Feature Importance Varies: The relative importance of formal features (REGISTER, REDUCTION, MORPHO-SYNTACTIC VARIANT) is not uniform across all sections. For example, morpho-syntactic features are more informative in narrative sections, while reduction features are more useful in informal sections.
- Conceptual Outliers: In semantically homogeneous sections, infrequent formal features can help identify conceptual outliers, suggesting a term might belong to a different semantic class than its peers.
Key Results
The study achieved varying F1-scores across different EHR sections, with the best performance observed in sections with a smaller number of target classes (e.g., Medication, Therapy, Examination).
- Overall Performance: The average F1-score across all sections was 0.69.
- Feature Importance: On average, MORPHO-SYNTACTIC features were the most important predictors, followed by REGISTER. REDUCTION was generally less informative.
- Section-Specific Performance:
- Medication: Achieved the highest F1-score (0.99).
- Examination: High F1-score (0.85).
- Therapy: High F1-score (0.86).
- Anamnesis: Good F1-score (0.73).
- Comments: Lower F1-score (0.64), but REDUCTION was a significant predictor here.
- Conclusion: Lower F1-score (0.48), with REGISTER being a strong predictor.
Limitations and Future Work
- Sample Size: The study focused on a small sample of highly frequent concepts. The informativeness of formal features for low-frequency concepts remains to be tested.
- Generalizability: The experiment was conducted on a specific clinical corpus (Belgian Dutch EHRs for diabetes patients). Future work should replicate the experiment on larger scales, with more diverse concept samples, and across different clinical specialties and institutions to test generalizability.
- Future Directions: The authors plan to expand the scale of the experiment and explore the method's applicability to a wider range of clinical data.
Summary updated