Using Large Language Models to Detect Anxiety and Nausea/Vomiting Documentation in Clinical Notes of Patients With...

Area of Science:

• Computational linguistics in healthcare
• Natural Language Processing (NLP) for clinical data analysis
• Oncology patient care and symptom management

Background:

• Large language models (LLMs) show promise for named entity recognition (NER) in healthcare.
• Accurate symptom detection in electronic health records (EHRs) is crucial for patient care.
• Identifying psychological and physical symptoms in cancer patients aids treatment planning.

Purpose of the Study:

• To evaluate LLM performance in identifying anxiety and nausea/vomiting symptoms in cancer patients' clinical notes.
• To compare fine-tuning versus prompt-based learning strategies for symptom detection.
• To assess the utility of LLM-based symptom detection for oncology care teams.

Main Methods:

• Analysis of clinical notes from 8,490 cancer patients using named entity recognition (NER).
• Pretraining Bio Clinical BERT and Bio GPT models on clinical text.
• Implementation of Symptom-BERT and Symptom-GPT frameworks using fine-tuning and prompt-based learning.
• Evaluation of model performance using F1 scores.

Main Results:

• Fine-tuning with Symptom-BERT achieved the highest F1 scores: 0.989 for nausea/vomiting and 0.912 for anxiety.
• Fine-tuning outperformed prompt-based learning, especially for well-documented physical symptoms.
• Anxiety was detected in 28.69% of patients, and nausea/vomiting in 39.31%.

Conclusions:

• Fine-tuning LLMs, particularly Symptom-BERT, offers superior accuracy for symptom detection in clinical narratives.
• LLM-based symptom detection can aid oncology nurses in early recognition and intervention.
• Automated symptom monitoring through LLMs enhances patient-centered care in oncology.