Negation Resolution

The foundational step of identifying lexical triggers that signal the absence of a clinical finding. Cues range from explicit determiners like 'no' and 'denies' to complex prepositional phrases such as 'without evidence of'.

• Rule-based methods (e.g., Negex) use regular expressions to match known trigger lists
• Contextual embedding models (e.g., NegBERT) disambiguate triggers from non-negating uses of the same vocabulary
• Pseudo-negation patterns like 'not only... but also' must be excluded to prevent false positives

Once a cue is detected, the system must determine the exact span of tokens whose meaning is inverted. Scope typically extends from the cue to the end of the clause or until a termination term is encountered.

• ConText algorithm extends Negex by defining scope termination rules using syntactic boundaries and conjunction analysis
• Dependency parsing identifies the grammatical subordinates of the cue to precisely map affected entities
• Span-level classification treats contiguous token sequences as single units for labeling, reducing boundary fragmentation errors

The binary or multi-class categorization of a clinical statement's factuality status. At minimum, this distinguishes affirmed (positive polarity) from negated (negative polarity) findings.

• Assertion status labels include Present, Absent, Uncertain, Historical, and Experiencer
• Double negation patterns like 'not unlikely' require logical resolution to avoid misclassification as negative
• Confidence scoring provides probabilistic outputs, enabling downstream thresholding for human review

Negation is not always absolute—findings may be negated only in the present while having occurred in the past. Temporal negation triggered by phrases like 'no longer present' or 'resolved' must be distinguished from permanent absence.

• Historical negation classifies conditions as part of the patient's past medical history rather than active problems
• ConText's temporal extension handles 'history of' triggers to prevent misattribution of resolved conditions
• Accurate temporal resolution is critical for constructing valid patient timelines and problem lists

Determining who the clinical finding applies to prevents false attribution of conditions to the patient. A finding may be affirmed but belong to a family member or contact.

• Experiencer negation triggered by phrases like 'mother has' or 'family history of' redirects the entity away from the patient
• ConText algorithm explicitly models experiencer as a distinct context alongside negation and temporality
• Failure to resolve experiencer leads to false positives in the patient's active problem list

Rigorous measurement ensures clinical safety. The cost of missed negation is incorrectly attributing a disease to a patient.

• Negation precision: proportion of correctly flagged negations among all flagged—minimizes false alarms
• False negative rate: proportion of actual negations missed—the critical safety metric where errors directly impact patient records
• BioScope corpus serves as the standard benchmark, containing annotated clinical free-text, biological papers, and abstracts for training and evaluation

The most straightforward pattern where a negation cue directly precedes the clinical finding.

Example: 'The patient denies any chest pain or shortness of breath.'

• Cue: 'denies'
• Scope: 'chest pain', 'shortness of breath'
• Resolved Assertion: Both findings are absent

Failure to resolve this negation would incorrectly populate the patient's problem list with two false conditions, potentially triggering unnecessary cardiac workups.

A negation cue word appears but does not semantically negate the target finding, requiring contextual disambiguation.

Example: 'The patient was treated not only for pneumonia but also for sepsis.'

• Trigger word: 'not'
• Actual meaning: Both pneumonia and sepsis are affirmed
• Pattern: 'not only X but also Y' is an emphatic affirmation

Rule-based systems like Negex can produce false positives here. Transformer models using contextual embeddings distinguish this pattern by attending to the full syntactic structure.

A finding that was true in the past but is no longer active at the time of documentation.

Example: 'History of hypertension, currently normotensive on medication.'

• Cue: 'History of'
• Temporal context: Past medical history
• Resolved Assertion: Hypertension is historical, not currently active

The ConText algorithm extends Negex to classify this as a distinct assertion status, preventing the condition from appearing as an active problem while preserving it in the patient's longitudinal record.

The clinical finding is affirmed but applies to a family member or contact, not the patient themselves.

Example: 'Patient's mother had breast cancer at age 45.'

• Experiencer: 'mother'
• Finding: 'breast cancer'
• Resolved Assertion: Finding is negated for the patient, attributed to family history

Without experiencer detection, this would incorrectly add breast cancer to the patient's active diagnosis list. ConText handles this by identifying the subject of the finding.

Two negation elements cancel each other, producing an affirmative statement that requires logical resolution.

Example: 'This is not an uncommon presentation for Lyme disease.'

• Cues: 'not', 'uncommon'
• Logical resolution: 'not uncommon' = common
• Resolved Assertion: The presentation is consistent with Lyme disease (affirmed)

This pattern is particularly challenging for token-level classifiers. Span-level classification approaches treat the entire construction as a unit to avoid misclassification.

A negation cue is combined with epistemic modality, expressing doubt about the absence rather than certainty.

Example: 'The patient reports no fever, but the reliability of this history is questionable.'

• Negation cue: 'no' (negates 'fever')
• Uncertainty cue: 'questionable reliability'
• Resolved Assertion: Fever is negated with low confidence

Advanced systems assign a confidence score to this extraction, flagging it for human review rather than treating it as a definitive absence. This is critical for clinical decision support where missing a fever could alter sepsis screening.

A specific word or phrase that triggers the semantic reversal of a clinical concept. Common examples include 'no', 'denies', 'without evidence of', and 'absence of'. The NegEx algorithm relies on a curated lexicon of these triggers, often paired with pseudo-negation filters to avoid false positives from phrases like 'not only' or 'not just'. Accurate cue identification is the critical first step in the resolution pipeline.

The specific span of tokens within a sentence whose meaning is inverted by a negation cue. Determining scope defines the boundary of what clinical concept is being ruled out. For example, in 'no acute fracture or dislocation', the scope extends over both findings. Scope resolution often uses syntactic dependency parsing or rule-based windowing to capture all affected entities without over-extending into unrelated clauses.

A linguistic construction containing a negation trigger word that does not actually negate a clinical condition. Examples include:

• 'not only pneumonia but also...'
• 'not otherwise specified'
• 'did not respond to treatment' Disambiguating these from true negation is essential to prevent false positives. Modern systems use contextual embeddings to distinguish pseudo-negation from genuine semantic reversal.

A classification label assigned to a clinical named entity indicating whether the concept is present, absent, or uncertain in the patient record. This forms the core output of factuality detection systems. The ConText algorithm extends basic negation to include additional statuses like historical and experienced_by_other, enabling precise patient timeline construction and preventing false attribution of conditions.

A linguistic pattern where two negation elements cancel each other out to form an affirmative statement. Examples include 'not unlikely' or 'cannot rule out'. These constructions require logical resolution to avoid misclassification—treating 'cannot rule out pneumonia' as negated would incorrectly suppress a finding the clinician is actively considering. Resolution systems must detect and invert double negation spans.

A transformer-based language model specifically fine-tuned on the BioScope corpus to perform token-level negation and speculation detection. Unlike rule-based systems like NegEx, NegBERT leverages contextual embeddings to disambiguate triggers based on surrounding words. It achieves state-of-the-art performance on negation resolution benchmarks by learning nuanced patterns that escape regular expression approaches.