ConText Algorithm

Determines if a clinical finding is explicitly stated as absent.

• Mechanism: Uses a trigger list (e.g., 'denies', 'without', 'no evidence of') and pseudo-grammar rules to define the scope of the negation.
• Example: 'Patient denies chest pain' → Negated.
• Differentiation: Unlike NegEx, ConText handles complex syntactic structures where the negation cue and the target are separated by multiple clauses.

Classifies a condition as occurring in the past (historical) or the present (recent).

• Mechanism: Identifies temporal triggers like 'history of', 'prior', 'last year' versus 'acute', 'new onset', 'current'.
• Example: 'Patient has a history of hypertension' → Historical.
• Clinical Impact: Prevents a resolved childhood condition from being incorrectly flagged as an active problem in a current encounter.

Identifies who the clinical finding refers to—the patient, a family member, or a donor.

• Mechanism: Detects possessive phrases and relational nouns like 'mother', 'father', 'sibling', or 'donor'.
• Example: 'Mother had breast cancer at 45' → Family member.
• Criticality: Essential for accurate family history extraction and preventing false diagnoses in the patient's active problem list.

Flags findings discussed in a conditional or future context that are not currently true.

• Mechanism: Triggers include 'if', 'consider', 'rule out', 'might be', 'differential includes'.
• Example: 'Will consider antibiotics if fever persists' → Hypothetical.
• Nuance: Distinguishes a working diagnosis under investigation from a confirmed, actionable finding, preventing premature clinical decision support alerts.

Distinguishes between statements about the patient and generic medical knowledge or instructions.

• Mechanism: Identifies educational boilerplate, general facts, or instructions not specific to the patient's current state.
• Example: 'Patients with diabetes should monitor glucose daily' → Generic subject.
• Application: Filters out noise from patient handouts or templated text that does not represent an actual finding for the individual.

Assesses the level of confidence associated with a clinical assertion.

• Mechanism: Uses a lexicon of hedging cues ('possible', 'suspected', 'cannot rule out') and definite cues ('confirmed', 'diagnosed with', 'positive for').
• Example: 'Suspected pulmonary embolism' → Uncertain.
• Integration: Often combined with negation detection to correctly interpret complex phrases like 'cannot rule out malignancy' as an uncertain positive, not a negation.

The foundational predecessor to ConText, NegEx uses regular expressions and trigger terms to determine whether a clinical finding is negated. It identifies negation cues like 'no evidence of' and defines a scope window—typically ending at a termination term or sentence boundary. While limited to binary negation detection, NegEx established the pattern-matching paradigm that ConText extended to handle temporality and experiencer properties.

The task of determining the exact span of text affected by a negation cue. For example, in 'no chest pain, but shortness of breath,' the scope must correctly exclude 'shortness of breath' from negation. Modern approaches use bidirectional LSTMs or transformer attention weights to learn scope boundaries, moving beyond the fixed-window heuristics of early ConText implementations.

The process of mapping relative clinical expressions to standardized, absolute time formats. Tools like HeidelTime and SUTime resolve ambiguous phrases such as 'q.d.' (once daily) or 'three weeks ago' into ISO-standard timestamps. ConText's temporality module—classifying conditions as recent, historical, or hypothetical—provides the semantic layer that complements these surface-level normalization systems.

The computational task of identifying which meaning of a polysemous or homonymous word is activated in context. In clinical text, this resolves ambiguities like 'MI' (myocardial infarction vs. mitral insufficiency) or 'cold' (temperature vs. viral illness). WSD provides the sense inventory that ConText then modifies—a negated 'MI' must first be correctly disambiguated before its contextual properties can be assigned.

A dynamic vector representation where a word's encoding changes based on surrounding text. ClinicalBERT and BioBERT generate these embeddings, allowing the same abbreviation to have different vector representations depending on whether it appears in a cardiology note or a dermatology report. ConText algorithms increasingly leverage these embeddings to replace rule-based context windows with attention-weighted semantic neighborhoods.

The task of grounding a recognized clinical mention to its unique, unambiguous identifier in a knowledge base like UMLS or SNOMED CT. After ConText determines that a condition is affirmed and recent, entity linking maps it to a Concept Unique Identifier (CUI). This two-stage pipeline—contextual modification followed by normalization—ensures that only clinically present findings populate structured problem lists and billing codes.