Semantic annotation of clinical events for generating a problem list

Area of Science:

• Clinical Informatics
• Natural Language Processing
• Medical Informatics

Background:

• Accurate clinical problem lists are crucial for patient care and medical research.
• Existing methods for extracting problem status from clinical text have limitations.
• Semantic understanding of clinical notes is essential for automated problem list generation.

Purpose of the Study:

• To develop and evaluate an annotation schema for clinical problems and their temporal attributes.
• To assess the human annotator performance using the proposed schema.
• To determine the contribution of semantic annotations in predicting the status of clinical problem mentions.

Main Methods:

• Developed an annotation schema for clinical problems, attributes, and temporal modifiers.
• Evaluated human annotator recall and agreement on clinical named entities and attributes.
• Utilized machine learning models (Support Vector Machine, Naïve Bayes, Decision Tree) to predict problem status based on annotations.

Main Results:

• Human annotators achieved low to moderate recall for clinical named entities and attributes.
• Certain attributes (Experiencer, Existence, Certainty) were more informative for status prediction than others.
• Support Vector Machine demonstrated superior performance in predicting problem status compared to Naïve Bayes and Decision Tree.

Conclusions:

• The developed annotation schema captures valuable semantic information for clinical problem status determination.
• Semantic annotations derived from clinical reports can significantly aid in generating accurate problem lists.
• Machine learning, particularly SVM, is effective for automated clinical problem status prediction using annotated data.