Extracting entities with attributes in clinical text via joint deep learning

Area of Science:

• Natural Language Processing (NLP)
• Medical Informatics
• Machine Learning

Background:

• Clinical entity and attribute extraction is crucial for medical NLP.
• Traditional pipeline methods for this task suffer from error propagation between sequential subtasks.
• Existing methods often struggle with simultaneous recognition and relation extraction.

Purpose of the Study:

• To propose a novel joint deep learning method for simultaneous clinical entity/attribute recognition and relation extraction.
• To address the limitations of sequential pipeline approaches in medical NLP.
• To develop a unified framework integrating state-of-the-art NLP techniques.

Main Methods:

• A unified deep learning framework integrating bidirectional long short-term memory with conditional random field and bidirectional long short-term memory.
• Simultaneous consideration of relation constraints and subtask weights.
• Comparative analysis against pipeline and other joint deep learning methods on English and Chinese corpora.

Main Results:

• The proposed joint method achieved superior performance on both entity recognition (F1: 74.46% English, 89.32% Chinese) and relation extraction (F1: 50.21% English, 88.13% Chinese).
• Outperformed existing pipeline and joint deep learning methods on both English and Chinese datasets.
• Demonstrated significant improvements in accuracy for both tasks.

Conclusions:

• The joint deep learning approach effectively improves both clinical entity recognition and relation extraction.
• The method shows promise for enhancing the analysis of clinical text in multiple languages.
• This unified framework offers a more robust solution for complex medical NLP tasks.