Improving Patient Safety Event Report Classification with Machine Learning and Contextual Text Representation.

Area of Science:

• Medical Informatics
• Natural Language Processing
• Patient Safety

Background:

• Medical errors cause significant patient harm and healthcare costs in the US.
• Accurate classification of patient safety event (PSE) reports is crucial for identifying preventative measures.
• Current methods using static natural language processing (NLP) have limitations in contextual understanding, impacting classification accuracy.

Purpose of the Study:

• To improve the accuracy of patient safety event (PSE) report classification.
• To leverage contextual text representations from neural NLP methods for enhanced classification performance.
• To identify limitations in existing PSE classification taxonomies.

Main Methods:

• Utilized contextual text representations derived from neural NLP models (e.g., Roberta-base).
• Employed machine learning classifiers, including support vector machines (SVM), trained on these contextual representations.
• Evaluated classifier performance using metrics such as accuracy and ROCAUC, and analyzed confusion matrices.

Main Results:

• Contextual text representations significantly improved PSE classifier performance compared to static text representations.
• The best-performing classifier, an SVM with Roberta-base contextual representation, achieved 0.75 accuracy and 0.94 ROCAUC.
• Analysis revealed deficiencies in the current PSE classification taxonomy, including multi-class and conceptually related event types.

Conclusions:

• Neural NLP-based contextual text representations offer a superior approach for classifying patient safety event (PSE) reports.
• Improved classification accuracy can streamline reclassification efforts and enhance healthcare reporting systems.
• The findings contribute to advancing patient safety through more effective analysis of adverse event data.