Four sentences, four meanings
| Sentence | Interpretation to preserve |
|---|---|
| No chest pain today | Symptom denied in a stated time frame |
| Chest pain last year | Historical information |
| Evaluate possible cause of chest pain | A question or assessment context |
| Parent had chest pain | Information about another person |
These are illustrative phrases, not medical assessments. Their point is that identical keywords can appear in different assertions.
Copy-forward creates another problem
A phrase can be repeated across notes without representing a newly observed event. Counting every mention as a new symptom episode can inflate apparent frequency.
The note date also does not necessarily equal the date of the event described.
Evaluate the output you actually need
If the task is identifying current symptoms, test current-symptom extraction rather than generic keyword detection. Review errors involving negation, historical mentions, family history, and uncertain statements separately.
An impressive overall accuracy score can hide failures in uncommon but important contexts. Preserve the text evidence and extraction rules where appropriate so reviewers can trace a coded result back to its source. A structured output is useful only when it retains the meaning needed for the clinical or research question.
