Annotating risk factors for heart disease in clinical narratives for diabetic patients

Area of Science:

• Biomedical Informatics
• Natural Language Processing
• Clinical Data Mining

Background:

• Identifying cardiovascular disease risk factors in clinical narratives is crucial for patient care and research.
• Existing methods often require extensive manual annotation, posing a significant time and resource burden.

Purpose of the Study:

• To develop and evaluate a "light" annotation paradigm for identifying cardiac artery disease risk factors in longitudinal electronic medical records.
• To create a high-quality, gold-standard corpus for a clinically relevant natural language processing task.

Main Methods:

• Utilized a "light" annotation strategy on 1304 longitudinal medical records from 296 patients.
• Employed majority voting to establish the gold standard annotations.
• Achieved high inter-annotator agreement (average > 0.95) with the gold standard.

Main Results:

• Generated document-level annotations for heart disease risk factors and their temporal presence.
• The resulting corpus demonstrated high reliability, supporting studies on risk factor progression.
• Participating systems in the 2014 i2b2/UTHealth shared task achieved a mean F1 score of 0.815.

Conclusions:

• A "light" annotation approach can yield reliable, high-quality clinical data for natural language processing tasks.
• The created corpus effectively supports research into the progression of heart disease risk factors.
• This methodology offers a balance between annotation efficiency and data quality for clinical research.