Machine learning and rule-based approaches to assertion classification.

Area of Science:

• Natural Language Processing
• Clinical Informatics
• Machine Learning

Background:

• Assertion classification is crucial for extracting patient health status from clinical text.
• Existing methods like NegEx have limitations in handling complex assertions.
• Developing robust assertion classification systems is essential for clinical decision support.

Purpose of the Study:

• To compare a rule-based approach (Extended NegEx) with a machine learning approach (Statistical Assertion Classifier) for assertion classification.
• To evaluate the generalizability and utility of both assertion classification methods across different clinical corpora.
• To assess the impact of lexical and syntactic context on assertion classification accuracy.

Main Methods:

• Two assertion classification systems, Extended NegEx (ENegEx) and Statistical Assertion Classifier (StAC), were developed and applied.
• Both systems identify medical problems as present, absent, uncertain, or associated with someone else.
• Performance was evaluated using precision, recall, and F-measure metrics.

Main Results:

• Both ENegEx and StAC demonstrated general applicability to new corpora.
• The machine learning-based StAC achieved comparable generality to the rule-based ENegEx.
• StAC models trained on discharge summaries were successfully applied to radiology reports.

Conclusions:

• Statistical Assertion Classifier (StAC) shows strong performance and generalizability, outperforming Extended NegEx (ENegEx).
• Machine learning models for assertion classification can be effectively applied across different clinical document types.
• Contextual word windows of +/- 4 words are most beneficial for StAC's assertion classification accuracy.