Categorized diagnoses and procedure records in an administrative database improved mortality prediction

Area of Science:

• Health Informatics
• Clinical Epidemiology
• Data Science in Healthcare

Background:

• Comorbidity measures are crucial for predicting mortality in administrative databases.
• Existing models often lack the granularity of detailed diagnostic and procedural information.

Purpose of the Study:

• To evaluate the impact of subcategorized secondary diagnoses and recorded procedures on mortality prediction model performance.
• To compare the predictive power of different comorbidity extraction methods using the Japanese Diagnosis Procedure Combination database.

Main Methods:

• Analysis of adult patients with specific primary diagnoses (e.g., myocardial infarction, pneumonia) from a 1-year administrative database.
• Construction of logistic regression models using Charlson and Elixhauser comorbidity indices, with variations in diagnostic subcategorization.
• Evaluation of the association between specific procedures (CT scans, oxygen, urinary catheters, vasopressors) and in-hospital mortality.

Main Results:

• Models incorporating all secondary diagnoses showed higher predictive accuracy (C-statistics: Charlson 0.717, Elixhauser 0.762) compared to limited subcategories.
• The comprehensive model identified misclassifications of complications and primary diagnoses as comorbidities.
• Four specific procedures were significantly associated with increased mortality risk.

Conclusions:

• Subcategorized diagnoses and recorded procedures significantly improve the accuracy of mortality prediction in administrative databases.
• These detailed features are valuable for enhancing predictive models and should be considered for implementation in other healthcare data systems.