Epistemic Modality

Epistemic modality operates on a gradient of commitment rather than a binary true/false distinction. Speakers position propositions along a continuum from absolute certainty to complete uncertainty.

• High certainty: 'The patient definitely has pneumonia'
• Medium certainty: 'The patient probably has pneumonia'
• Low certainty: 'The patient might have pneumonia'
• Speculative: 'Could this be pneumonia?'

This gradient is what makes uncertainty detection a multi-class classification problem rather than simple polarity detection. Systems must learn to map lexical cues like possible, likely, and suspected to probability thresholds.

A subcategory of epistemic modality that encodes the source of information supporting a claim. In clinical text, evidentiality reveals whether a finding is based on direct observation, reported symptoms, or inference.

• Direct evidence: 'Chest X-ray shows consolidation'
• Reported evidence: 'Patient reports shortness of breath'
• Inferred evidence: 'Findings suggestive of infection'
• Hearsay: 'Per family, patient apparently fell'

Evidentiality markers help downstream systems weight the reliability of extracted information. A finding grounded in imaging carries different evidentiary value than a patient's subjective report.

A critical distinction in clinical NLP: epistemic stance is the writer's expressed attitude toward a proposition, while epistemic status is the actual veridical state of the finding in the patient record.

• Stance: 'I suspect the patient has diabetes' (clinician's belief)
• Status: The patient either has diabetes or does not (ground truth)

Uncertainty detection systems classify stance, not status. The phrase 'cannot rule out malignancy' expresses low certainty stance, but the actual status remains unknown. This gap is why human-in-the-loop review remains essential for high-stakes clinical decisions.

English modal verbs are the primary grammatical mechanism for expressing epistemic modality. Each carries distinct certainty implications that uncertainty detection systems must disambiguate.

• Must: High certainty inference ('This must be the source of infection')
• Should/ought to: Expectation based on evidence ('Labs should normalize')
• May/might/could: Possibility without commitment ('May represent artifact')
• Can't/couldn't: Strong negative inference ('Cannot be cardiac in origin')

The same modal can express different types of modality depending on context. Must can be epistemic (certainty) or deontic (obligation), requiring contextual disambiguation.

Epistemic modality markers have scope—the specific proposition or clause whose certainty is being modified. Identifying scope boundaries is essential for accurate clinical extraction.

• Narrow scope: 'Possible [pneumonia] in the [right lower lobe]'
• Wide scope: '[The patient has pneumonia] is likely'
• Embedded scope: 'CT shows [possible [nodule] in [left upper lobe]]'

Scope resolution errors can cause certainty leakage, where a hedge intended for one finding incorrectly modifies another. This is the same linguistic challenge addressed by the ConText algorithm for negation scope.

Two opposing epistemic strategies that modify the force of a clinical assertion:

Hedging reduces commitment and opens diagnostic possibilities:

• 'Mild irregularity may reflect early changes'
• 'Findings are nonspecific but could represent infection'

Boosting strengthens commitment and closes diagnostic uncertainty:

• 'Findings are classic for pneumonia'
• 'This definitively rules out malignancy'

Clinical text often contains mixed epistemic strategies within a single sentence, requiring models to resolve conflicting certainty signals. 'Cannot rule out' is a hedge despite containing the booster-like 'cannot.'

The most common clinical manifestation of epistemic modality, where a clinician hedges commitment to a finding.

• Example: 'The lesion is suspicious for malignancy'
• Example: 'Findings suggestive of pneumonia'
• Cue phrases: 'consistent with', 'compatible with', 'raises the possibility of'

These statements express inferential probability rather than definitive assertion, requiring downstream NLP systems to assign an 'uncertain' factuality status rather than treating the condition as confirmed.

Clinicians routinely list multiple possibilities ranked by likelihood, creating a layered epistemic landscape.

• Example: 'Chest pain likely musculoskeletal, but cannot rule out pulmonary embolism'
• Example: 'Probable** urinary tract infection versus possible nephrolithiasis'

This construction combines graded certainty ('likely', 'probable') with explicit uncertainty ('cannot rule out', 'versus'). Accurate extraction requires preserving the relative confidence levels rather than flattening all mentions to equal status.

Epistemic modality is often grounded in the source and quality of evidence supporting a claim, not just the claim itself.

• Example: 'Per patient report, symptoms began three days ago'
• Example: 'No objective evidence of infection on examination'
• Example: 'Radiographically confirmed left lower lobe consolidation'

The modality shifts from uncertain ('per report') to negated ('no evidence') to affirmed ('confirmed'), all within the same document. Systems must track the evidential basis to correctly weight extracted findings.

Epistemic modality intersects with time when clinicians express uncertainty about when or if a condition will manifest.

• Example: 'Patient at risk for developing sepsis'
• Example: 'Potential for future arrhythmia given family history'
• Example: 'Condition may progress to respiratory failure'

These statements express future-oriented possibility rather than present uncertainty. They require distinct classification from current speculative findings, as they represent risk prediction rather than diagnostic ambiguity.

Clinicians often express high certainty about a finding contingent upon a specific condition being met.

• Example: 'If cultures are positive, diagnosis is pyelonephritis'
• Example: 'Assuming no complications, discharge will be tomorrow'
• Example: 'In the event of desaturation, immediate intubation indicated'

The modality is hypothetical but definitive within its conditional frame. NLP systems must recognize that the certainty is real but scoped to a conditional trigger, preventing both false affirmation and false uncertainty classification.

Clinical language employs a spectrum of epistemic strength that goes beyond binary certain/uncertain classification.

• High certainty: 'definitively', 'unequivocally', 'classic presentation of'
• Moderate certainty: 'probably', 'likely', 'most consistent with'
• Low certainty: 'possibly', 'conceivably', 'cannot exclude'
• Indeterminate: 'unclear whether', 'indeterminate', 'equivocal'

Fine-grained epistemic classification enables downstream systems to prioritize confirmed findings for billing while flagging low-certainty mentions for clinical review.

A sub-task of uncertainty detection focused on linguistic devices that weaken a speaker's commitment to a proposition.

• Examples: 'It is suggestive of malignancy', 'The mass is likely benign'.
• Function: In clinical text, hedging allows clinicians to communicate differential diagnoses without asserting a definitive fact.
• Challenge: Requires models to understand subtle gradations of probability expressed through adverbial phrases.

The veridical assessment of a clinical event's occurrence, representing the output of a full epistemic modality analysis.

• Categories: Typically includes affirmed, negated, uncertain, and historical.
• Importance: Constructing an accurate patient timeline requires knowing not just what was mentioned, but if it actually happened.
• Relation: Epistemic modality provides the linguistic signals that determine the 'uncertain' category.

A classification label assigned to a clinical named entity indicating its presence, absence, or uncertainty in the patient record.

• Core Output: The primary deliverable of factuality detection systems like ConText.
• Values: 'Present', 'Absent', 'Possible', 'Hypothetical', 'Historical', 'Not Patient'.
• Application: Directly populates structured data fields in downstream analytics to prevent false diagnoses from speculative text.

A probabilistic output associated with an uncertainty prediction, quantifying the model's own certainty about the epistemic status.

• Utility: Allows clinical systems to threshold results or prioritize ambiguous cases for human-in-the-loop review.
• Example: A 0.95 probability of 'uncertain' vs. a 0.51 probability triggers different downstream actions.
• Calibration: Well-calibrated scores are critical for safe automation in high-stakes medical environments.