A method based on multi-standard active learning to recognize entities in electronic medical record.

Area of Science:

• Computer Science
• Medical Informatics

Background:

• Deep neural networks (DNNs) excel at Named Entity Recognition (NER) but require extensive annotated data.
• Electronic Medical Records (EMRs) present unique challenges due to specialized medical terminology and the difficulty/cost of expert annotation.

Purpose of the Study:

• To develop an efficient NER method for EMRs that addresses data annotation challenges.
• To reduce the dependency on large volumes of labeled data in specialized domains.

Main Methods:

• Proposed a multi-standard active learning strategy for EMR Named Entity Recognition (NER).
• Utilized criteria including labeled data volume, annotation cost, and data sampling balance to guide active learning.
• Introduced novel uncertainty calculation and measurement rules for neural network models in NER.
• Employed incremental training to accelerate iterative training within the active learning process.
• Developed an improved TF-IDF method incorporating Word2Vec for enhanced text vectorization.

Main Results:

• Achieved comparable NER accuracy in breast clinical EMRs with reduced data labeling.
• Demonstrated a 66.67% reduction in required labeled data compared to traditional random sampling supervised learning.
• The proposed vectorization method considers word frequency for improved text representation.

Conclusions:

• The multi-standard active learning method is effective for NER in specialized domains like EMRs.
• This approach significantly lowers the annotation burden for medical text data.
• The study highlights the potential for more efficient and accurate information extraction from EMRs.