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

Extracting structured information from unstructured clinical text remains a critical challenge in healthcare. This study introduces a robust framework for clinical assertion detection, integrating domain-specific embeddings like BioBERT, contextualized learning, and a performance-driven voting mechanism. By leveraging pre-trained models, including BiLSTM-CNN-Char architectures, the framework effectively classifies clinical assertions across diverse dimensions, such as Polarity, Subject, and Tense. The result demonstrates high F1-scores (0.95-0.98) across key categories, highlighting its adaptability to complex clinical contexts. The framework outperforms existing approaches by addressing the challenges of nested concepts, small datasets, and imbalanced data. By incorporating multiple models and consolidating outputs through voting, the framework also reduces dependency on any single model, enhancing reliability and adaptability. These findings emphasize its potential for enhancing clinical decision-making, improving patient care, and supporting scalable healthcare research initiatives.