A flexible data-driven comorbidity feature extraction framework

Area of Science:

• Medical Informatics
• Machine Learning
• Health Services Research

Background:

• Disease diagnostic codes are crucial for patient outcome prediction.
• Current methods for utilizing diagnostic codes in machine learning have limitations in dimensionality and predictive power.

Purpose of the Study:

• To propose a novel clustering-based feature extraction methodology for disease diagnostic codes.
• To enhance the predictive accuracy of patient severity of condition and readmission risk using machine learning.

Main Methods:

• Developed a clustering algorithm using co-occurrence statistics to identify disease clusters and reduce data dimensionality.
• Optimized cluster generation and used clusters as features for predictive modeling.
• Utilized the National Inpatient Sample (NIS) dataset and Electronic Health Records (EHR) for model development and testing.

Main Results:

• The proposed cluster-based feature set demonstrated significant improvements in predictive accuracy compared to traditional comorbidity frameworks.
• Achieved gains of 10.7-22.1% in predicting congestive heart failure (CHF) severity.
• Showed improvements of 4.65-5.75% in predicting diabetes readmission risk.

Conclusions:

• The novel clustering-based feature extraction methodology effectively utilizes disease diagnostic information for improved patient outcome prediction.
• This approach offers a more powerful alternative to existing comorbidity indices for machine learning applications in healthcare.
• The framework shows promise for enhancing clinical decision support and risk stratification in diverse patient populations.