A Performance-Based Voting Framework for Assertion Detection in Clinical Notes.

Area of Science:

• Natural Language Processing in Healthcare
• Clinical Informatics

Background:

• Extracting structured information from unstructured clinical text is a significant challenge.
• Existing methods struggle with complex clinical data, including nested concepts and imbalanced datasets.

Purpose of the Study:

• To develop a robust framework for clinical assertion detection.
• To improve the accuracy and reliability of extracting structured information from clinical text.

Main Methods:

• Integration of domain-specific embeddings (BioBERT) and contextualized learning.
• Utilizing BiLSTM-CNN-Char architectures and a performance-driven voting mechanism.
• Leveraging pre-trained models for classification of clinical assertions (Polarity, Subject, Tense).

Main Results:

• Achieved high F1-scores (0.95-0.98) across key assertion categories.
• Demonstrated superior performance compared to existing approaches, particularly with challenging datasets.
• Framework shows adaptability to complex clinical contexts and data limitations.

Conclusions:

• The proposed framework enhances clinical decision-making and patient care.
• It offers a reliable and adaptable solution for scalable healthcare research.
• The voting mechanism reduces dependency on single models, increasing robustness.