Diagnosis code assignment: models and evaluation metrics.

Area of Science:

• Medical Informatics
• Natural Language Processing
• Health Data Science

Background:

• Healthcare data volume is rapidly increasing due to health information technology adoption.
• Accurate automated coding of clinical data is crucial for data analysis and research.
• Discharge summaries contain rich information for automated diagnosis code assignment.

Purpose of the Study:

• To develop and evaluate automated methods for assigning International Classification of Diseases, 9th Revision (ICD9) codes from clinical discharge summaries.
• To compare a novel hierarchy-based classification approach against a traditional flat classifier.
• To introduce new evaluation metrics that account for the hierarchical structure of ICD9 codes.

Main Methods:

• Utilized the Multiparameter Intelligent Monitoring in Intensive Care II (MIMIC II) dataset, including discharge summaries and ICD9 codes.
• Implemented and compared two classification approaches: a flat classifier and a hierarchy-based classifier.
• Developed novel evaluation metrics assessing code similarity and hierarchical accuracy.

Main Results:

• The hierarchy-based classifier achieved a higher F-measure (39.5%) compared to the flat classifier (27.6%).
• Novel metrics demonstrated the hierarchy-based approach's superior ability to identify correct ICD9 sub-trees.
• Error analysis suggested potential underestimation of performance due to imperfections in gold-standard codes.

Conclusions:

• Hierarchy-based classification is more effective for automated ICD9 coding from MIMIC patient discharge summaries.
• Automated ICD9 coding is a suitable task for data and tool sharing to advance research.
• Collaborative efforts using shared resources can significantly improve the state-of-the-art in clinical coding.