Negex Algorithm

Negex relies on a curated lexicon of negation cues—words or phrases that semantically reverse a clinical finding. The algorithm scans sentences for these triggers.

• Core Triggers: 'no', 'denies', 'without', 'absence of', 'free of'
• Pseudo-Negation Handling: Phrases like 'not only' or 'not just' are explicitly excluded to prevent false positives
• Phrasal Triggers: Multi-word expressions such as 'no evidence of' or 'ruled out' are treated as single units

The trigger list is domain-adapted for clinical text, distinguishing between general negation and medical-specific phrasing.

Once a negation cue is identified, Negex applies grammatical heuristics to define the negation scope—the span of text whose meaning is inverted.

• Forward Window: The algorithm looks up to 6 tokens to the right of the trigger
• Termination Conditions: The scope ends at a conjunction ('but', 'however'), a punctuation mark, or another clinical assertion
• Prepositional Boundaries: Phrases like 'except for' or 'but' act as hard scope terminators

This rule-based scoping prevents over-negation of unrelated clinical concepts in complex sentences.

Negex operates on UMLS-tagged clinical entities rather than raw text. The algorithm requires pre-identified medical concepts as input.

• Input Format: Sentences with pre-annotated clinical terms from the Unified Medical Language System
• Output: Each concept receives an assertion status of 'affirmed' or 'negated'
• Integration Pattern: Typically deployed as a post-processing step after a Medical NER pipeline

This design separates entity recognition from factuality classification, allowing Negex to be swapped or upgraded independently.

A critical feature of Negex is its explicit handling of pseudo-negation—constructions where negation words do not actually negate a condition.

• Double Negation: 'not unlikely' is recognized as an affirmative statement
• Emphatic Constructions: 'no significant change' does not negate the existence of a finding
• Exclusion Lists: Predefined patterns like 'not only... but also' are filtered before trigger matching

Without this disambiguation, systems would incorrectly flag conditions like 'no significant pneumonia' as absent when the condition is actually present but mild.

Negex has well-documented constraints that inform its appropriate use cases.

• No Uncertainty Handling: Cannot detect hedged or speculative language ('possible pneumonia', 'suspected MI')
• No Historical Context: Does not distinguish between current negation and past conditions ('no longer has diabetes')
• No Experiencer Detection: Cannot identify when a finding applies to a family member rather than the patient
• Fixed Window: The 6-token scope window may miss long-distance dependencies in complex sentences

These limitations motivated the development of the ConText algorithm as a direct extension.

Negex achieves strong performance on standard clinical negation corpora despite its rule-based simplicity.

• Precision: ~95% on radiology reports and discharge summaries
• Recall: ~85-90%, with false negatives primarily from complex syntactic structures
• F1 Score: Consistently above 0.90 on the BioScope corpus negation subtask
• Speed: Processes thousands of documents per second with minimal compute overhead

This performance profile makes Negex ideal for high-throughput clinical data pipelines where deterministic behavior is preferred over black-box models.

The specific span of tokens in a sentence whose meaning is inverted by a negation cue. Defining the scope is the critical second step after trigger detection. For example, in 'no evidence of acute chest pain or shortness of breath', the scope extends from 'acute' to 'breath'. Incorrect scope resolution is a primary source of error in rule-based systems, as the algorithm must accurately determine which clinical entities fall within the negated window and which remain affirmed.

Linguistic constructions that contain a negation trigger word but do not semantically negate a clinical condition. These are common sources of false positives in Negex. Key examples include:

• Double negation: 'not unlikely'
• Affirmative phrases: 'not only diabetes but also...'
• Idiomatic usage: 'did not hesitate to diagnose' Disambiguation often requires syntactic parsing or contextual embeddings to distinguish true negation from these pseudo-negation patterns.

The classification label assigned to a clinical named entity indicating its factuality in the patient record. This is the core output of all negation and uncertainty detection systems. Standard categories include:

• Present: The condition is affirmed for the patient.
• Absent: The condition is explicitly ruled out (Negex output).
• Uncertain: The condition is speculated or hedged.
• Historical: The condition occurred in the past but is not active.
• Family/Other: The condition applies to someone other than the patient. These labels are critical for accurate patient timeline construction and cohort identification.