Ontology-driven weak supervision for clinical entity classification in electronic health records

Area of Science:

• Natural Language Processing
• Machine Learning
• Health Informatics

Background:

• Extracting entities and temporality from clinical notes is crucial for health record analysis.
• Manual data labeling for clinical entity tasks is time-consuming and hindered by privacy concerns.
• The COVID-19 pandemic emphasized the need for rapid machine learning model training for clinical notes.

Purpose of the Study:

• To present Trove, a novel framework for weakly supervised entity classification in clinical notes.
• To offer an agile, shareable, and modifiable alternative to manual data labeling.
• To validate Trove's performance against benchmark tasks and real-world patient data.

Main Methods:

• Developed a weakly supervised learning framework (Trove) utilizing medical ontologies and expert-generated rules.
• Applied Trove for entity classification in clinical text.
• Validated the framework on six benchmark tasks and patient records from Stanford Health Care.

Main Results:

• Trove achieved performance comparable to models trained on manually labeled data.
• The framework demonstrated effectiveness in analyzing electronic health records for COVID-19 related symptoms and risk factors.
• Trove provides an easily shareable and modifiable approach to clinical entity extraction.

Conclusions:

• Trove offers a viable and efficient solution for weakly supervised entity classification in clinical notes.
• The framework addresses key challenges in data creation and privacy for machine learning in healthcare.
• Trove has the potential to accelerate clinical text analysis for various health informatics applications.