Web-Based Application Based on Human-in-the-Loop Deep Learning for Deidentifying Free-Text Data in Electronic Medical...

Area of Science:

• Health Informatics
• Medical Data Privacy
• Natural Language Processing

Background:

• Electronic medical records (EMRs) contain valuable clinical data for research.
• Protecting patient privacy necessitates deidentifying personally identifiable information (PII) in EMR free text.
• Manual deidentification is inefficient, driving the need for automated solutions.

Purpose of the Study:

• Develop an accurate, user-friendly, web-based system for deidentifying EMR free text, named DEFT.
• Enhance adoption in real-world settings through features like interactive learning and collaboration.
• Facilitate secondary use of clinical data while maintaining patient privacy.

Main Methods:

• DEFT utilizes a Bidirectional Long Short-Term Memory-Conditional Random Field (BiLSTM-CRF) deep learning model with RoBERTa embeddings.
• An interactive learning loop with preannotation accelerates manual deidentification.
• The system supports project management, customizable PII types, and user access control.

Main Results:

• DEFT achieved superior performance on the 2014 i2b2 dataset, with microaverage F1-scores of 0.9627.
• Real-world application on clinical notes yielded a microaverage F1-score of 0.9507.
• Preannotation increased manual annotation efficiency by 43%.

Conclusions:

• DEFT provides an accessible solution for deidentifying EMR free text for researchers and data custodians.
• The system's interactive learning loop and user-friendly interface lower the technical barrier for deidentification tasks.
• DEFT facilitates secure and efficient utilization of clinical narrative data.