Leveraging Natural Language Processing for Symptom Identification in Acute Myeloid Leukemia Using Clinical Notes from Electronic Health Records.

The study used data from one Midwestern academic hospital, so results may not apply to other hospitals, EHR systems, or AML patient groups. The paper gives different encounter counts: 3046 in the extraction description and 2742 in the note-composition section. The NLP tool can misread negated phrases; the paper gives an example where a note denying chest pain was still flagged as chest pain. Copy-paste documentation can inflate symptom counts and introduce redundancy. The validation sample had 240 notes, with physicians 58% and registered nurses 0.4%, versus 41.5% and 7.3% in the full note set. Symptom terms that could fit more than one category were assigned to one category, which may limit how symptoms appear across groups. The very large note-level sample can make small documentation differences statistically significant without clear clinical relevance. The study did not test whether NLP extraction improves symptom management or patient outcomes.

The authors have no conflicts of interest to disclose.

Do not read the high symptom rates as proof patients experienced those symptoms at those rates. The system counted terms in notes, and negation errors and copy-paste documentation can inflate counts.