Efficient Active Learning for Electronic Medical Record De-identification.

Area of Science:

• Health Informatics
• Machine Learning
• Natural Language Processing

Background:

• Electronic medical records (EMRs) require de-identification for secondary data use.
• De-identifying unstructured clinical text is challenging and annotation-intensive.
• Current methods demand significant human annotation resources.

Purpose of the Study:

• To integrate active learning into the EMR de-identification workflow.
• To reduce the need for extensive human annotation in de-identification.
• To improve the efficiency of training de-identification models.

Main Methods:

• Incorporated an active learning framework into the de-identification process.
• Applied the active learning approach to clinical trial and i2b2 datasets.
• Compared active learning performance against traditional passive learning methods.

Main Results:

• Active learning requires less training data to achieve high performance.
• An F-measure of 0.9 was reached with only 40 documents using active learning (batch size 10).
• Passive learning required at least 25% more data for comparable model performance.

Conclusions:

• Active learning enhances de-identification model training efficiency.
• This approach reduces the burden of human annotation in clinical data processing.
• Active learning offers a more data-efficient strategy for EMR de-identification.