Clinical concept annotation with contextual word embedding in active transfer learning environment.

Area of Science:

• Natural Language Processing
• Medical Informatics
• Machine Learning

Background:

• Unstructured clinical data presents challenges for automated information extraction.
• Accurate identification of clinical concepts (Problem, Treatment, Test) is crucial for healthcare applications.

Purpose of the Study:

• To develop and evaluate an active learning approach for automated clinical concept extraction.
• To classify clinical concepts with high precision and recall from unstructured electronic health records.

Main Methods:

• Utilized a lexical-based approach for initial data labeling to support active learning.
• Employed contextual word embedding similarity with BERT variants (ClinicalBERT, DistilBERT, SCIBERT) for concept classification.
• Trained deep learning models, including Convolutional Neural Networks (CNNs), and large language models (LLMs) using active learning.

Main Results:

• SCIBERT demonstrated strong performance in active transfer learning, achieving 73.97% F1-score.
• CNNs trained with various embeddings (BERTBase, DistilBERT, SCIBERT, ClinicalBERT) reached 89-93% testing accuracy.
• ClinicalBERT as an LLM achieved the highest performance with 96% testing accuracy.

Conclusions:

• The integrated active learning methodology with SCIBERT and CNNs enhances clinical concept extraction.
• The approach improves annotation efficiency and maintains high accuracy, showing significant potential for clinical applications.