Clinical Safety Incident Taxonomy Performance on C4.5 Decision Tree and Random Forest.

Area of Science:

• Health Informatics
• Artificial Intelligence
• Patient Safety

Background:

• Clinical safety incident (CSI) classification is crucial for patient safety.
• Automated classification systems can improve accuracy and efficiency in managing CSI reports.
• Existing classification systems may require refinement for optimal performance.

Purpose of the Study:

• To evaluate artificial intelligence (AI) classifiers for automated CSI classification.
• To compare the effectiveness of the Generic Reference Model (GRM) and World Health Organization (WHO) patient safety classifications.
• To develop an improved taxonomy for enhanced CSI classification accuracy.

Main Methods:

• Applied C4.5 decision tree (DT) and random forest (RF) classifiers.
• Trained classifiers on 3600 CSIs from an Incident Information Management System (IIMS).
• Utilized a bag-of-words approach with GRM and WHO taxonomies, and an improved WHO (WHO-I) taxonomy.

Main Results:

• The RF classifier outperformed the DT classifier.
• The WHO taxonomy demonstrated superior performance compared to the GRM taxonomy.
• The enhanced WHO-I taxonomy further improved classification accuracy, particularly for previously poorly performing classes.

Conclusions:

• AI, specifically the RF classifier, is effective for automated CSI classification.
• The WHO taxonomy, especially when enhanced (WHO-I), is more suitable for automated CSI reporting than the GRM.
• Refining existing taxonomies can significantly improve the performance of AI-driven patient safety systems.