Phenonizer: A Fine-Grained Phenotypic Named Entity Recognizer for Chinese Clinical Texts

Area of Science:

• Natural Language Processing
• Clinical Informatics
• Biomedical Data Analysis

Background:

• Electronic medical records contain valuable clinical information in free text.
• Current biomedical named entity recognition (BioNER) methods often use coarse annotations, missing nuanced details like symptom negation.
• Accurate extraction of phenotypic entities is crucial for understanding patient characteristics, especially in diseases like COVID-19.

Purpose of the Study:

• To develop a fine-grained BioNER system for Chinese clinical texts.
• To address the challenge of recognizing negated phenotypic entities.
• To improve the extraction of structured clinical data for better patient profiling.

Main Methods:

• Development of the Human-machine Cooperative Phenotypic Spectrum Annotation System (HCPSAS) and a fine-grained Chinese clinical corpus.
• Proposal of Phenonizer, a BioNER model utilizing BERT for contextual representation, bidirectional LSTM for local features, and CRF for label sequencing.
• Comparison of Phenonizer against Word2Vec-based methods and evaluation on different dataset granularities.

Main Results:

• Phenonizer achieved an F1-score of 0.896 on a COVID-19 dataset, outperforming Word2Vec methods.
• Character embeddings trained on clinical corpora improved the F1-score by 0.0103.
• Fine-grained datasets slightly boosted F1-scores (approx. 0.005) and enabled distinction between negated and presented symptoms.
• Phenonizer demonstrated strong generalization performance with an F1-score of 0.8389.

Conclusions:

• Phenonizer offers a feasible and effective approach for extracting symptom information from Chinese clinical texts.
• Fine-grained annotation and advanced deep learning models are key to improving BioNER accuracy and handling negation.
• The developed system and corpus contribute to advancing clinical data analysis and patient profiling.