Expert guided natural language processing using one-class classification.

Area of Science:

• Biomedical Informatics
• Natural Language Processing
• Machine Learning

Background:

• Automated phenotype identification from clinical notes is crucial for data reuse.
• This study explores expert-guided feature selection combined with one-class classification for text processing.

Purpose of the Study:

• Compare one-class classification to binary classification.
• Evaluate expert-selected text snippets for feature utility.
• Assess model robustness against irrelevant text.

Main Methods:

• Trained one-class support vector machines (1C-SVMs) and two-class SVMs (2C-SVMs) to detect breast cancer mentions.
• Utilized manually annotated notes (88 positive, 88 negative) for comparison.
• Evaluated models on balanced and imbalanced datasets (10,000 records, 1.4% prevalence).

Main Results:

• On balanced data, 1C-SVMs with snippets matched 2C-SVMs on whole notes (F=0.92).
• On imbalanced data, 1C-SVMs significantly outperformed 2C-SVMs (F=0.61 vs. F=0.17), driven by improved precision.

Conclusions:

• One-class SVMs trained on expert-selected text sections excel over traditional binary classifiers on imbalanced clinical data.
• This approach enhances the identification of low-prevalence phenotypes in real-world datasets.