Semi-supervised classification of patient safety event reports

Area of Science:

• Health Informatics
• Natural Language Processing
• Machine Learning

Background:

• The Veterans Health Administration (VHA) patient safety reporting system handles over 100,000 reports annually.
• These critical reports are predominantly composed of unstructured natural language text.
• Efficiently extracting actionable information from this large volume of text is a significant challenge.

Purpose of the Study:

• To develop and evaluate a semi-supervised machine learning method for enhanced information extraction from VHA patient safety reports.
• To improve the efficiency and accuracy of categorizing and analyzing patient safety data.
• To extend the capabilities of existing patient safety taxonomies.

Main Methods:

• A semi-supervised classification approach was employed, utilizing both labeled and unlabeled patient safety reports.
• The method predicts categories for unlabeled reports by leveraging patterns learned from labeled data.
• Two distinct scenarios of the semi-supervised learning process were investigated.

Main Results:

• Both examined scenarios demonstrated strong predictive performance.
• The semi-supervised method proved effective in accurately assigning categories to patient safety reports.
• The approach showed good results in handling reports with varying and overlapping topics.

Conclusions:

• The proposed semi-supervised method shows significant promise for assisting researchers and analysts in processing patient safety reports.
• This technique enables faster and more accurate separation of reports based on their content.
• The method effectively utilizes narrative data within reports to enhance and expand static patient safety taxonomies.