Uncertainty Cue

These cues express the degree of certainty regarding a proposition. They form the core of speculation detection in clinical text.

• High Certainty: 'probable', 'likely', 'consistent with'
• Low Certainty: 'possible', 'cannot rule out', 'unlikely'
• Direct Speculation: 'suspected', 'concerning for', 'presumed'

These terms directly modify the factuality status of an extracted medical concept, requiring classification systems to distinguish between a confirmed finding and a differential diagnosis.

Hedging cues weaken the speaker's commitment without explicitly stating uncertainty. They often indicate the source or strength of evidence.

• Evidential Hedges: 'appears to show', 'suggests', 'may represent'
• Clinician Stance: 'I suspect', 'the patient reports', 'per outside records'

These constructions are critical for experiencer detection and distinguishing a clinician's objective finding from a patient's subjective report, preventing false attribution of conditions.

These cues place a medical finding within a non-factual context, often indicating a future or contingent event rather than a current diagnosis.

• Conditional: 'if symptoms persist', 'should the patient develop'
• Hypothetical: 'to rule out', 'versus', 'differential includes'
• Recommendation: 'consider', 'evaluate for', 'monitor for'

The ConText algorithm specifically targets these triggers to classify findings as hypothetical, ensuring they are not recorded as active problems in the patient's structured record.

The scope of an uncertainty cue defines the span of text whose factuality is modified. Incorrect scope resolution is a primary source of error.

• Preceding Scope: 'possible pneumonia' (cue before target)
• Following Scope: 'infiltrate cannot be excluded' (cue after target)
• Discontinuous Scope: 'mass is possibly malignant' (cue interrupts target phrase)

Modern span-level classification models using contextual embeddings from architectures like NegBERT are designed to resolve these complex syntactic boundaries with high precision.

Not every instance of a trigger word signals genuine clinical doubt. Pseudo-uncertainty requires contextual disambiguation to prevent false positives.

• Rhetorical Use: 'Could it be anything else?' (not modifying a specific finding)
• Historical Certainty: 'The patient likely had this as a child.' (certain about the past)
• Negated Uncertainty: 'There is no question of malignancy.' (double negation resolving to certainty)

Failure to resolve these patterns leads to a high false negative rate for affirmed findings, incorrectly labeling confirmed diagnoses as speculative.

Uncertainty detection systems output a probabilistic confidence score for each classification. This enables nuanced downstream handling.

• High Confidence (>0.95): Automatic structured data mapping
• Medium Confidence (0.7-0.95): Flagged for human-in-the-loop review
• Low Confidence (<0.7): Excluded from automated pipelines, routed for expert adjudication

This tiered approach balances negation precision with recall, ensuring that ambiguous hedging like 'cannot entirely exclude' is reviewed by a clinician before affecting the patient's problem list.

Auxiliary verbs that express possibility, probability, or necessity rather than factuality.

• may represent early-stage disease
• could indicate an underlying infection
• might be consistent with pneumonia
• should resolve with treatment

These verbs weaken the assertion of a diagnosis, requiring downstream systems to flag the associated finding as speculative rather than confirmed.

Adverbs that reduce the force of a clinical assertion by introducing doubt or approximation.

• possibly related to medication side effects
• likely represents a benign finding
• probably secondary to dehydration
• presumably viral in origin

These modifiers often appear before a diagnosis and signal that the clinician is working with incomplete information or differential reasoning.

Adjectives that frame a diagnosis as tentative or under investigation rather than definitive.

• suspected pulmonary embolism
• presumed community-acquired pneumonia
• probable acute coronary syndrome
• possible deep vein thrombosis
• apparent treatment failure

These terms are common in admission notes and emergency department documentation where definitive testing is still pending.

Multi-word expressions that indicate a finding is based on limited or indirect evidence.

• cannot rule out myocardial infarction
• suggestive of an infectious process
• concerning for malignancy
• raises the possibility of endocarditis
• differential diagnosis includes

These phrases are critical for clinical decision support systems to recognize, as they indicate active diagnostic deliberation rather than confirmed pathology.

Expressions that limit the certainty of a finding to a specific time window or indicate pending resolution.

• appears to be resolving
• currently thought to be
• at this time, etiology unclear
• pending further workup
• awaiting confirmatory testing

These cues are essential for constructing accurate patient timelines and distinguishing between active, resolving, and indeterminate clinical states.

Conditional or subjunctive structures that place a finding in a hypothetical rather than actual context.

• if infection is present, antibiotics indicated
• in the event of recurrence
• should symptoms persist
• consider appendicitis if pain migrates
• would be consistent with sarcoidosis

These constructions require scope detection to ensure that the hypothetical condition is not extracted as an active diagnosis in structured data.