Secure Extraction of Personal Information from EHR by Federated Machine Learning

Area of Science:

• Health Informatics
• Artificial Intelligence
• Data Privacy

Background:

• Extracting Personally Identifiable Information (PII) from Electronic Health Records (EHRs) poses significant privacy and security risks.
• Existing methods often struggle to balance data utility with robust confidentiality measures.

Purpose of the Study:

• To evaluate the efficacy of Federated Learning (FL) for secure PII extraction from French EHRs.
• To compare the performance of FL models against individual, non-federated models.

Main Methods:

• A simulation involving 20 hospitals using a multilingual BERT model within a Federated Learning framework.
• Comparison of federated models trained collaboratively across hospitals versus individual models trained solely on local data.

Main Results:

• Federated Learning models demonstrated superior performance compared to individual models.
• The Global FL model achieved an F1 score of 75.7%, closely approaching the performance of a centralized approach (78.5%).
• FL successfully preserved data confidentiality throughout the PII extraction process.

Conclusions:

• Federated Learning offers a promising solution for secure and effective PII extraction from sensitive EHR data.
• FL facilitates collaborative health data analysis while mitigating privacy concerns.
• The study advocates for wider adoption of FL in health data research and applications.