Estimation of a predictor's importance by Random Forests when there is missing data: risk prediction in liver surgery...

Area of Science:

• Hepatobiliary Surgery
• Medical Informatics
• Statistical Modeling

Background:

• Liver surgery has advanced, but extended resections still carry risks.
• Preoperative laboratory parameters may predict postoperative complications.
• Missing data is common in clinical datasets, complicating analysis.

Purpose of the Study:

• To investigate the impact of missing data handling on the predictive relevance of preoperative laboratory parameters in liver surgery.
• To compare complete case analysis, Random Forests, and multiple imputation for analyzing liver surgery data with missing values.
• To identify key laboratory predictors for liver failure and postoperative complications.

Main Methods:

• Analysis of a large liver surgery database.
• Comparison of complete case analysis, Random Forests (handling missing values internally), and multiple imputation.
• Extensive simulation study to evaluate the performance of different missing data approaches.
• Assessment of variable importance measures.

Main Results:

• Complete case analysis distorts variable importance.
• Random Forests with internal missing value handling appropriately reflect decreased variable relevance.
• Multiple imputation can reveal potential complete-data variable relevance if imputation is successful.
• Lactate and bilirubin show potential association with liver failure and complications.

Conclusions:

• The choice of method for handling missing data significantly impacts the assessment of predictor relevance in liver surgery.
• Complete case analysis should be avoided due to distorted results.
• Random Forests offer a robust approach for analyzing datasets with missing values.
• Lactate and bilirubin warrant further investigation as predictors of adverse outcomes in liver surgery.