Symptom-BERT: Enhancing Cancer Symptom Detection in EHR Clinical Notes

Area of Science:

• Artificial Intelligence in Medicine
• Natural Language Processing for Healthcare
• Clinical Informatics

Background:

• Extracting cancer symptom data from clinical narratives is crucial for personalized symptom prediction and management.
• Advanced language models can significantly improve the detection of symptom data within clinical notes.

Purpose of the Study:

• To develop and evaluate a pretrained language model for detecting and extracting cancer symptoms from clinical notes.
• To assess the performance of the Symptom-BERT model in identifying various cancer symptom groups.

Main Methods:

• Pretrained a Bio-Clinical BERT model on one million unlabeled clinical documents.
• Fine-tuned Symptom-BERT on 1112 annotated notes for 13 cancer symptom groups.
• Validated externally using 180 synthetic notes generated by ChatGPT-4.

Main Results:

• Symptom-BERT achieved high internal performance (F1-score 0.933, AUC 0.929) and strong external validation (AUC 0.834).
• The model demonstrated higher accuracy in detecting physical symptoms (e.g., Pruritus) compared to psychological symptoms (e.g., anxiety).

Conclusions:

• Specialized pretraining on domain-specific data significantly enhances language model performance for medical applications.
• The Symptom-BERT model represents a significant advancement in AI for patient-centered cancer symptom detection and management.