Semantic classification of diseases in discharge summaries using a context-aware rule-based classifier

Area of Science:

• Medical Informatics
• Natural Language Processing

Background:

• Automated classification of clinical discharge summaries is crucial for medical research.
• Extracting semantic labels (present, absent, questionable, unmentioned) enhances data utility.

Purpose of the Study:

• To present a context-aware, rule-based semantic classification technique for clinical discharge summaries.
• To accurately assign semantic labels to disease mentions within clinical texts.

Main Methods:

• A rule-based system preprocesses text by removing misleading parts.
• Text is segmented into positive, negative, and uncertain contexts.
• A sequence of binary classifiers assigns semantic labels.

Main Results:

• The system achieved high performance on the i2b2 Obesity Challenge.
• Results included F(1)-macro = 0.80 for textual and F(1)-macro = 0.67 for intuitive classification.
• The system secured first and second place in the challenge's subtasks.

Conclusions:

• A simple rule-based classifier can surpass machine learning for semantic classification with limited or sparse data.
• This approach facilitates more accurate data extraction from clinical notes for research.