Automated mutual exclusion rules discovery for structured observational codes in echocardiography reporting

Area of Science:

• Medical Informatics
• Artificial Intelligence in Medicine
• Cardiology Reporting

Background:

• Structured reporting in medicine aims to improve data processing and communication.
• Structured echocardiography reports can contain contradictory finding codes (FCs), reducing quality and reliability.
• Manual creation of rules to detect these contradictions is labor-intensive.

Purpose of the Study:

• To develop a machine learning approach for automatically discovering mutual exclusion rules between FCs in structured echocardiography reports.
• To automate the identification of previously unknown rules for detecting report contradictions.

Main Methods:

• A machine learning model was trained on 101,211 structured echocardiography reports.
• Semantic and statistical analysis was used to identify patterns of mutual exclusion between FCs.
• Ground truth for rule validation was derived from an existing, prospectively evaluated rule set.

Main Results:

• The machine learning approach achieved an F-measure of 0.439 and an AUC of 0.885 on an unseen test set.
• The method demonstrated the potential to support the manual creation of rules for detecting contradictions.
• Expert review confirmed the discovery of previously unknown mutual exclusion rules.

Conclusions:

• The proposed machine learning approach can effectively discover mutual exclusion rules in echocardiography reports.
• This automation can significantly aid in the creation and refinement of rule sets for ensuring data quality in structured medical reports.
• The findings suggest a pathway to enhance the reliability of machine-assisted processing of medical information.