A patient-centered approach to developing and validating a natural language processing model for extracting...

Area of Science:

• Natural Language Processing (NLP)
• Computational Linguistics
• Health Informatics

Background:

• Patient-reported symptoms are crucial for healthcare quality but challenging to capture.
• Existing NLP models primarily focus on clinical notes, not patient-generated text.
• There's a need for NLP tools tailored to patient narratives.

Purpose of the Study:

• To develop and validate a novel NLP model for extracting patient-reported symptoms from pharmaceutical care records.
• To evaluate the model's performance on diverse, patient-generated narratives.
• To compare the new model against existing tools for clinical text analysis.

Main Methods:

• Developed a transformer-based named entity recognition (BERT-CRF) model.
• Utilized pharmaceutical care records ('Subjective' sections) from patients on anticancer drugs.
• Created ground-truth data using two annotation guidelines.
• Validated the model on external patient-generated blog data.

Main Results:

• The BERT-CRF model significantly outperformed an existing Japanese clinical text tool.
• Achieved an F1 score > 0.8 on pharmaceutical care records.
• Extracted > 98% of physical symptom entries from patient blogs, a 20% improvement.

Conclusions:

• Fine-tuning NLP models with patient-specific narrative data is essential.
• The developed model effectively captures nuanced and colloquial symptom expressions.
• This approach enhances the analysis of patient-generated health data.