An active learning pipeline to automatically identify candidate terms for a CDSS ontology-measures, experiments, and...

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Natural Language Processing for Ontologies

Background:

• Clinical Decision Support System (CDSS) ontologies are crucial for standardizing medical vocabulary and data integration.
• Traditional ontology development is manual, time-consuming, and requires extensive domain expertise.
• Automating key phrase identification for CDSS ontologies can significantly improve efficiency.

Purpose of the Study:

• To present an active learning framework for automatic identification of key phrases relevant to CDSS ontologies.
• To reduce the manual labor and domain expertise required for ontology development and long-term maintenance.
• To improve the accuracy and efficiency of CDSS ontology creation through automated methods.

Main Methods:

• A BiLSTM CRF model integrated into a human-in-the-loop active learning pipeline.
• Uncertainty sampling strategy for document selection, prioritizing low-confidence model instances for human review.
• Novel uncertainty aggregation methods (KPSum, KPAvg, DOCSum, DOCAvg) combined with uncertainty measures (MTP, TE, Margin) for document-level confidence scoring.

Main Results:

• The proposed active learning framework effectively identifies key phrases for CDSS ontologies.
• New uncertainty aggregation methods enhance the document selection process in active learning.
• The approach leads to more transparent, replicable, and efficient prioritization of documents for annotation.

Conclusions:

• Active learning significantly facilitates CDSS ontology development by reducing manual effort.
• The framework supports human experts, particularly in the long-term maintenance of ontologies.
• Automated key phrase identification is a valuable strategy for efficient and accurate ontology engineering.