Unified concept and assertion detection using contextual multi-task learning in a clinical decision support system.

Area of Science:

• Natural Language Processing
• Medical Informatics
• Machine Learning

Background:

• Assertions like negation and speculation modify clinical findings in Electronic Health Records (EHRs).
• Accurate assertion detection is crucial for effective clinical decision support systems.
• Existing methods using biomedical embeddings often fail to leverage inter-task knowledge transfer.

Purpose of the Study:

• To propose a novel neural model for unified concept and assertion detection in clinical notes.
• To integrate task-specific fine-tuning and multi-task learning within a coherent framework.
• To enhance the performance of concept and assertion detection through inter-task knowledge transfer.

Main Methods:

• Developed a novel neural model employing multi-task learning and task-specific fine-tuning.
• Utilized a unified framework that leverages the hierarchical relationship between concept and assertion detection tasks.
• Evaluated the model on real-world clinical notes datasets (n2c2 2010, n2c2 2012, NegEx).

Main Results:

• Achieved significant performance improvements in concept detection (+1.69 F1 on n2c2 2010, +2.96 F1 on n2c2 2012).
• Demonstrated enhanced assertion recognition (+2.89 F1 and +3.77 F1 on datasets).
• Showcased substantial gains in negation detection under low-resource settings (+2.4 F1) and improved speculation detection (+2.09 F1).

Conclusions:

• The proposed unified multi-task learning framework effectively transfers knowledge between related tasks.
• This approach significantly enhances the accuracy of concept and assertion detection in clinical text.
• Represents the first contextual multi-task system for unified detection in clinical decision support applications.