Negation detection, uncertainty modelling, and assertion status — why standard NER is insufficient

Every extracted clinical entity carries an assertion status: present, absent, or uncertain. Named entity recognition answers a different question — it tells you that a clinical concept was mentioned, not what the author asserted about it. Coding only the present set therefore requires that the absent and uncertain sets be identified with equal confidence, because the cost of misclassifying them is asymmetric and high.

In a typical clinical note a large fraction of mentioned conditions are explicitly negated or hedged. Differential diagnoses are enumerated in order to be excluded. Review of systems is largely a list of things the patient does not have. An extraction pipeline that treats mention as assertion will produce a code set several times larger than the correct one, weighted toward exactly the conditions the clinician ruled out.

Assertion status also interacts with subject. A condition can be present, absent, or uncertain for the patient, or it can belong to a family member, or it can be hypothetical — a condition described in patient instructions as something to watch for. Subject attribution is a separate axis and needs to be modelled separately.

Negation cues frequently sit outside the entity span and can be separated from it by several clauses. A single cue can scope over a coordinated list, so the negation applies to entities many tokens away. Conversely, a cue can be scoped narrowly by an intervening conjunction, so an adjacent entity is not negated at all. Local span features cannot distinguish these cases because the discriminating information is syntactic structure, not proximity.

Rule-based scope algorithms handle the common patterns well and are still the right baseline, but they degrade on the fragmented syntax that dictation produces. When sentence boundaries are missing, a scope window defined in tokens will run past the end of the clinical statement it was meant to cover and negate entities from the next one.

The practical implication is that assertion classification needs sentence-level or clause-level structure as input, and that structure has to be recovered from noisy text rather than assumed. Systems that skip this step tend to show good aggregate negation accuracy driven by the easy majority of simple cases, and poor accuracy on the compound and coordinated constructions where the errors actually matter.

Hedged language — possible, cannot rule out, suspect, consistent with, versus — marks a genuinely distinct state. Coding rules treat uncertain diagnoses differently depending on the setting: for inpatient records a diagnosis documented as probable or suspected at discharge may be coded as if established, while in the outpatient setting it may not be. The same phrase therefore produces different correct behaviour depending on encounter type.

That setting-dependence means uncertainty cannot be collapsed into a confidence score. Confidence describes the system's belief about its own extraction; uncertainty describes the clinician's assertion. Conflating the two loses the information the coding rule needs, and produces a system that cannot correctly implement inpatient versus outpatient policy.

Degree of hedging also matters for routing. Weakly hedged findings in an outpatient note are usually not codeable and can be suppressed; strongly asserted findings hedged only by a single qualifier are frequently worth a reviewer's attention. Emitting the cue and its strength alongside the entity is what makes that routing possible.

A fourth axis sits alongside assertion and subject: time. A condition can be current, historical, or resolved, and the coding treatment differs in each case. Historical conditions may be codeable as personal history codes rather than as active diagnoses; resolved conditions frequently are not codeable at all for the current encounter.

Temporal signals come from explicit date expressions, from tense, and from section placement — and these three sources routinely disagree in copy-forward documentation, where a paragraph written months ago appears in the present tense in today's note. Resolving that conflict requires treating section placement as evidence rather than as ground truth.

In audit practice, current-versus-historical errors are as common as negation errors and are harder for reviewers to catch, because the extracted entity and the assigned code both look correct in isolation. Only comparison against the encounter timeline reveals the error.

False-positive codes from ruled-out findings are the highest-risk error class in an audit, because they are indefensible against the source document. There is no argument to be made: the note explicitly says the patient does not have the condition. Where such codes affect risk adjustment or DRG assignment, the exposure is not limited to the claim.

For that reason assertion status deserves its own measurement and its own acceptance threshold, reported separately from overall coding accuracy. Precision on the present class and recall on the absent class are the two figures that matter, and a system should be evaluated on notes selected specifically for dense negation and hedging rather than on a random sample.

The review policy should follow the same logic. Any code whose supporting extraction carried an uncertain or contested assertion status is a candidate for human review regardless of terminal confidence, and the reviewer interface should show the cue and its scope so the decision takes seconds rather than requiring a full re-read of the note.