Missing data in medical databases: impute, delete or classify?

Area of Science:

• Medical Informatics
• Data Science in Healthcare
• Clinical Data Management

Background:

• Intensive care unit (ICU) databases often contain missing data due to multiple information sources.
• Missing data negatively impacts the performance of predictive risk modeling and medical guideline development.
• Traditional methods like imputation or deletion introduce bias and reduce statistical power.

Purpose of the Study:

• To present a new approach for managing missing data in ICU databases.
• To enhance the overall performance of predictive risk modeling.

Main Methods:

• Utilized a statistical classifier followed by fuzzy modeling to identify data for imputation.
• Developed a simulation test bed for performance evaluation.
• Validated the approach on a previously published ICU database.

Main Results:

• The new method significantly improved classification accuracy by 11%, sensitivity by 13%, and specificity by 10%.
• Area under the receiver-operator curve (AUC) improved by up to 13%.
• Successfully managed datasets with missing data ranging from 10-50%.

Conclusions:

• The proposed method effectively improves predictive modeling performance in simulated and real-world ICU data.
• This advanced missing data management technique offers a valuable tool for researchers aiming to enhance ICU predictive risk modeling.