Two-stage Federated Phenotyping and Patient Representation Learning.

Area of Science:

• Medical Informatics
• Natural Language Processing
• Machine Learning

Background:

• Electronic medical records contain vast unstructured clinical text data.
• Manual information extraction from clinical notes is inefficient.
• Existing natural language processing (NLP) models lack generalizability across institutions.

Purpose of the Study:

• To develop a novel federated NLP approach for analyzing distributed clinical notes.
• To improve the performance of clinical NLP tasks by leveraging multi-institutional data.
• To facilitate knowledge discovery within a learning health system.

Main Methods:

• A two-stage federated natural language processing (NLP) method was developed.
• The approach allows analysis of clinical notes from different healthcare providers without data centralization.
• Performance was evaluated using obesity and comorbidities phenotyping.

Main Results:

• The federated NLP method demonstrated effective utilization of distributed clinical notes.
• The approach improved the quality of the specific clinical task (phenotyping).
• This represents the first application of federated machine learning in clinical NLP.

Conclusions:

• Federated NLP offers a scalable solution for extracting insights from heterogeneous clinical text data.
• This method enhances the generalizability of clinical NLP models.
• It supports collaborative knowledge advancement across the healthcare system.