Leveraging Natural Language Processing for Symptom Identification in Acute Myeloid Leukemia Using Clinical Notes from Electronic Health Records.
Source: PubMed, NCBI / U.S. National Library of Medicine
Patients with acute myeloid leukemia (AML) experience severe, co-occurring, and fluctuating symptoms during treatment. Accurate identification is critical but often limited by under documentation and unstructured electronic health record notes. To develop and validate a natural language processing (NLP) system to extract symptoms from inpatient clinical notes, characterize prevalence and co-occurrence of documented symptoms, and examine whether documentation patterns vary by patient- and note-level factors. We analyzed 78,392 clinical notes from 812 AML patients admitted between 2006 and 2021. Ten symptom categories were defined (pain, gastrointestinal, myelosuppression, cardiopulmonary, skin, fatigue, anxiety/anger, central nervous system, depression, and sleep). The NLP system was validated against 240 manually annotated notes, with performance assessed by precision, recall, and F1 score. Exploratory analyses using generalized estimating equations estimated odds of documentation by sex, age, and author type. The NLP system achieved high performance across all symptom categories (average F1 = 0.90). Gastrointestinal (97.3%), pain (95.9%), and myelosuppression (95.7%) were most frequently documented, with extensive co-occurrence across encounters. Fatigue and depression were less common. Men had lower odds of depression documentation than women, and older patients had lower odds across multiple domains. Compared with physicians, advanced practice registered nurses more often
Abstract
Patients with acute myeloid leukemia (AML) experience severe, co-occurring, and fluctuating symptoms during treatment. Accurate identification is critical but often limited by under documentation and unstructured electronic health record notes. To develop and validate a natural language processing (NLP) system to extract symptoms from inpatient clinical notes, characterize prevalence and co-occurrence of documented symptoms, and examine whether documentation patterns vary by patient- and note-level factors. We analyzed 78,392 clinical notes from 812 AML patients admitted between 2006 and 2021. Ten symptom categories were defined (pain, gastrointestinal, myelosuppression, cardiopulmonary, skin, fatigue, anxiety/anger, central nervous system, depression, and sleep). The NLP system was validated against 240 manually annotated notes, with performance assessed by precision, recall, and F1 score. Exploratory analyses using generalized estimating equations estimated odds of documentation by sex, age, and author type. The NLP system achieved high performance across all symptom categories (average F1 = 0.90). Gastrointestinal (97.3%), pain (95.9%), and myelosuppression (95.7%) were most frequently documented, with extensive co-occurrence across encounters. Fatigue and depression were less common. Men had lower odds of depression documentation than women, and older patients had lower odds across multiple domains. Compared with physicians, advanced practice registered nurses more often documented cardiopulmonary symptoms, while other provider groups documented fewer symptoms overall. NLP enables accurate, scalable extraction of symptom data from unstructured notes, supporting large-scale surveillance and predictive modeling in AML. Findings highlight the need for standardized documentation and tailored interventions to address symptom risks across patient groups.