Natural Language Processing Accurately Differentiates Cancer Symptom Information in Electronic Health Record...

Area of Science:

• Computational linguistics
• Medical informatics
• Oncology

Background:

• Identifying cancer symptoms in electronic health records (EHR) is crucial for patient care.
• Existing natural language processing (NLP) systems require improvement for detecting diverse symptoms and distinguishing them from negated symptoms or medication side effects.

Purpose of the Study:

• To evaluate the accuracy of NLP in detecting 14 distinct symptom groups within cancer patient EHR narratives.
• To assess NLP's ability to differentiate observed symptoms from negated symptoms and medication-related side effects.

Main Methods:

• Extracted 902,508 notes from 11,784 cancer patients.
• Developed a gold standard corpus of 1,112 notes for symptom labeling.
• Trained an embeddings-augmented NLP system integrating human and machine intelligence.

Main Results:

• The embeddings-augmented NLP model achieved a high F1 score of 0.877.
• Accuracy varied by symptom, with pruritus (F1=0.937) and swelling (F1=0.787) being the highest and lowest, respectively.
• 41% of notes contained symptom documentation, with pain being most frequent (29%) and impaired memory least frequent (0.7%).

Conclusions:

• NLP is feasible for detecting multiple cancer symptom groups in EHR narratives.
• The developed embeddings-augmented NLP system surpasses conventional machine learning in symptom detection and differentiation.