Modeling major lung resection outcomes using classification trees and multiple imputation techniques.

Area of Science:

• Thoracic Surgery
• Medical Statistics
• Health Informatics

Background:

• Operative risk modeling for major lung resection is often imprecise due to incomplete covariate data.
• Missing values in clinical databases can introduce bias and reduce statistical power.
• Traditional methods like case deletion lead to loss of valuable information and precision.

Purpose of the Study:

• To evaluate the utility of imputation techniques for handling missing predictor variables in major lung resection risk modeling.
• To identify key predictors of pulmonary, cardiovascular, overall complications, and mortality after major lung resection using imputation.

Main Methods:

• Analysis of major lung resection patients (1980-2006) for complication and mortality predictors.
• Classification and Regression Tree (CART) methods to determine initial predictive variables.
• Multiple imputation techniques applied to variables with missing data (e.g., serum albumin, DLCO%).
• Logistic regression models fitted using CART variables and imputed covariates.

Main Results:

• Serum albumin (32% missing) and diffusing capacity (DLCO%, 13% missing) had substantial missing values.
• Significant predictors identified included DLCO%, FEV1%, performance status, serum albumin, FEV1/FVC ratio, resection extent, operation year, and age.
• Serum albumin emerged as a strong predictor for overall complications.

Conclusions:

• Imputation techniques enhance the identification of significant predictive variables in surgical risk models, even with incomplete data.
• Serum albumin is a crucial, previously under-reported predictor of overall complications following major lung surgery.
• Utilizing imputation in clinical databases can improve the accuracy and efficiency of surgical risk assessment.