Word Sense Disambiguation (WSD)

The initial retrieval step that queries a sense inventory—typically the UMLS Metathesaurus—to fetch all possible meanings of an ambiguous abbreviation. For 'MI', this returns concepts like Myocardial Infarction (C0027051), Mitral Insufficiency (C0026266), and Mental Institution (C0025344).

• Relies on pre-compiled abbreviation lexicons and knowledge base lookups
• Generates a candidate set that downstream scoring models will rank
• Must handle out-of-vocabulary abbreviations not present in static inventories

Transforms the ambiguous token and its surrounding text into a dynamic vector representation using transformer architectures like ClinicalBERT. Unlike static embeddings, these vectors shift based on context—'MI' in a cardiology note receives a different embedding than 'MI' in a dermatology note.

• Captures bidirectional context from both left and right of the target token
• Encodes semantic relationships between the abbreviation and surrounding clinical entities
• Enables cosine similarity comparison against candidate sense embeddings

Constrains the candidate sense space by filtering on high-level UMLS Semantic Types. If the context indicates a medication discussion, all non-drug candidates are pruned before scoring. This dramatically reduces the search space and prevents nonsensical mappings.

• Distinguishes between 'Clinical Drug', 'Disease or Syndrome', and 'Laboratory Procedure'
• Uses section header awareness to infer expected semantic types
• Reduces false positives by eliminating category-inappropriate candidates

Leverages the self-attention weights from transformer layers to quantify which context words most influence the disambiguation decision. For 'CHF' in 'CHF exacerbation with pulmonary edema,' the model attends heavily to 'pulmonary edema' to select Congestive Heart Failure over 'Congenital Hepatic Fibrosis.'

• Provides interpretable evidence for why a specific sense was selected
• Enables confusion pair analysis to identify systematically ambiguous cases
• Attention heatmaps serve as audit trails for clinical reviewers

Extends disambiguation beyond the immediate sentence by incorporating global document signals such as the patient's problem list, encounter diagnosis codes, and SOAP note section headers. An abbreviation in the 'Assessment' section carries different priors than one in 'Family History.'

• Uses section header awareness to weight candidate senses by clinical context
• Integrates longitudinal patient data for persistent abbreviation resolution
• Resolves locally ambiguous cases where the sentence alone is insufficient

The final stage that maps the resolved surface form to a unique, unambiguous identifier in a target terminology. After disambiguating 'MI' to myocardial infarction, the system normalizes it to SNOMED CT 22298006 and optionally cross-maps to ICD-10-CM I21.9 for billing workflows.

• Produces standardized concept IDs for downstream FHIR resource mapping
• Handles lexical variants—'heart attack,' 'MI,' and 'myocardial infarction' all map to the same concept
• Enables interoperability across EHR, billing, and analytics systems

A dynamic vector representation of a word that changes based on surrounding text, enabling models like BERT to distinguish between the cardiological and dermatological senses of 'MI'.

• Static vs. Dynamic: Unlike Word2Vec, contextual embeddings generate different vectors for the same word in different sentences
• Mechanism: Transformer attention layers weigh surrounding tokens to produce a context-aware representation
• Clinical Application: A contextual embedding for 'MI' in 'chest pain, rule out MI' will be closer to 'myocardial infarction' in vector space than to 'mitral insufficiency'

The process of mapping a shortened clinical form like 'CHF' to its full intended meaning, such as 'Congestive Heart Failure,' using surrounding context to select the correct expansion from a sense inventory.

• Sense Inventory: A pre-compiled list of possible expansions, often sourced from UMLS Metathesaurus entries
• Scoring Mechanism: Candidate expansions are ranked by semantic similarity between the context and each candidate's definition
• Critical Distinction: Expansion is the surface-form mapping; disambiguation is the underlying sense selection

A family of transformer-based language models, including BioBERT and ClinicalBERT, pre-trained or fine-tuned on clinical corpora like MIMIC-III to capture domain-specific context for disambiguation tasks.

• Pre-training Data: Clinical notes, discharge summaries, and scientific literature provide the domain grounding
• Architecture: Bidirectional transformer encoders that generate contextualized token embeddings
• Performance Advantage: Clinical BERT variants consistently outperform general-domain BERT on abbreviation disambiguation benchmarks by 5-15% F1

A comprehensive compendium of biomedical vocabularies and standards, providing the Concept Unique Identifiers (CUIs) and semantic types used as a sense inventory for normalizing ambiguous mentions.

• Metathesaurus: Contains over 4 million concepts from 200+ source vocabularies, serving as the master sense inventory
• Semantic Network: 133 semantic types and 54 relationships that constrain candidate meanings
• CUI Grounding: The target for entity linking after disambiguation—'MI' resolved to C0027051 (Myocardial Infarction)

A disambiguation technique that constrains candidate meanings based on high-level UMLS categories, such as distinguishing a 'Procedure' from a 'Clinical Drug' when resolving an ambiguous acronym.

• Constraint Logic: If context indicates a medication, filter out candidates with semantic types like 'Disease or Syndrome'
• Type Hierarchy: Leverages the UMLS Semantic Network's 'is-a' relationships for broader or narrower filtering
• Error Reduction: Semantic type filtering eliminates 30-40% of incorrect candidates before fine-grained scoring

The task of grounding a recognized clinical mention to its unique, unambiguous identifier in a knowledge base like UMLS or SNOMED CT, a critical step following abbreviation resolution.

• Pipeline Position: Entity linking occurs after mention detection and disambiguation, producing a normalized concept ID
• Candidate Ranking: Uses cosine similarity between mention context embeddings and candidate concept embeddings
• Downstream Impact: Accurate entity linking enables cohort identification, clinical decision support, and ICD-10-CM code assignment