ConText Algorithm

Unlike Negex, which only determines if a finding is negated or affirmed, ConText classifies findings into multiple contextual dimensions:

• Historical: Conditions that occurred in the past but are not active ('history of asthma')
• Hypothetical: Findings mentioned in conditional or speculative scenarios ('if the patient develops fever')
• Experiencer: Determining whether the finding applies to the patient or a family member ('mother had breast cancer')
• Negated: Explicitly ruled out findings ('no evidence of pneumonia')

ConText relies on a curated dictionary of trigger terms that signal specific contextual categories:

• Negation triggers: 'no', 'denies', 'without', 'absence of'
• Historical triggers: 'history of', 'prior', 'previous', 'resolved'
• Hypothetical triggers: 'if', 'consider', 'rule out', 'possible'
• Experiencer triggers: 'mother', 'father', 'sibling', 'family history'

Each trigger is associated with a direction (forward/backward) and a termination rule that defines its scope boundary.

ConText defines the scope of a trigger using configurable token windows rather than full syntactic parsing:

• A trigger's influence extends a fixed number of tokens in its specified direction
• Scope terminates at pseudo-triggers like 'but', 'however', or 'except'
• Termination also occurs at sentence boundaries or section breaks
• This lightweight approach avoids the computational cost of full dependency parsing while maintaining high accuracy on clinical narratives

ConText explicitly accounts for pseudo-negation patterns where negation words do not actually negate a clinical finding:

• Double negation: 'not unlikely' resolves to affirmative
• Emphatic constructions: 'not only diabetes but also hypertension'
• Temporal non-negation: 'not yet diagnosed' implies future possibility

These rules prevent false positive negation classifications that would incorrectly remove valid clinical findings from structured data.

ConText operates as a post-processing step in clinical NLP pipelines:

1. NER System identifies clinical entities (diseases, symptoms, procedures)
2. ConText scans the surrounding text for trigger terms
3. Scope rules determine which entities fall within each trigger's influence
4. Assertion labels are assigned: Present, Absent, Historical, Hypothetical, or Family

This modular design allows ConText to enhance any named entity recognition system without retraining.

ConText achieves strong performance on clinical factuality benchmarks:

• Negation detection: F1 scores exceeding 0.90 on radiology reports and discharge summaries
• Historical classification: Accurately distinguishes past from active conditions, critical for problem list generation
• Experiencer detection: Prevents erroneous attribution of family history to the patient
• Limitations: Rule-based approach may miss novel or implicit negation patterns that deep learning models like NegBERT can capture through contextual embeddings

The output label assigned to each clinical named entity indicating its factuality in the patient record. Standard categories include:

• Present: The finding is affirmed for the patient
• Absent: The finding is explicitly negated
• Uncertain: The finding is hedged or speculative
• Historical: The finding occurred in the past but is not active
• Family History: The finding applies to a relative, not the patient This classification is the core deliverable of ConText and similar systems.

The process of determining the exact token span whose meaning is inverted by a negation cue. For example, in 'no evidence of acute pneumonia or pleural effusion', the scope must correctly extend from 'acute pneumonia' through 'pleural effusion'. Scope boundary detection is the most challenging aspect of negation resolution, as incorrect boundaries lead to false negatives where conditions are missed or false positives where affirmed findings are incorrectly negated.

A critical edge case where negation trigger words appear but do not actually negate a clinical finding. Examples include:

• 'not only pneumonia but also sepsis'
• 'did not improve but rather worsened'
• 'cannot exclude the possibility of' ConText must distinguish these affirmative constructions from true negation to prevent false positives. This requires syntactic pattern recognition beyond simple trigger matching.

Two specialized negation types that ConText uniquely handles beyond Negex:

Historical Negation: The condition occurred in the past but is not active now. Triggered by phrases like 'history of', 'resolved', or 'status post'.

Experiencer Negation: The finding is present but applies to a family member or contact, not the patient. Triggered by 'mother', 'father', 'sibling', etc.

Both prevent false attribution of conditions to the patient's active problem list.