Prediction of hospital mortality by classification trees and smoothing splines in patients with haematological malignancies admitted to the ICU

Introduction

Classification trees have not been amply used in an ICU setting. Logistic regression modeling for prognosis in the ICU is still the current standard. This study examines the possibility of the clinical use of classification trees for mortality prediction in patients with haematological malignancies admitted to the ICU. Recently, new modeling tools based on classification tree algorithms have become available [1, 2]. These new modeling tools have proven to be accurate and robust in domain applications outside the clinical setting. Moreover, classification tree models can be constructed easily by the ICU clinician and have the advantages over logistic regression to give a clear insight into the data.

Methods

Three hundred and fifty-two patients with haematological malignancies that were admitted to the ICU between 1997 and 2006 for a life-threatening complication were included. Two hundred and fifty-two patient records were used for training of the models and 100 were used for validation. In a first model 12 input variables were included, and in a second more complex model 17 input variables were used for model development. These two models were consecutively constructed using Classification and Regression Tree (CART^®), Multivariate Adaptive Regression Splines (MARS^®), Random Forests (RF™) and TREENET™ software (Salford Systems, San Diego, CA, USA). Discrimination was evaluated using the area under the receiver operating characteristic curves (± SE). All results were then compared with the results obtained by a logistic regression model.

Results

The area under the receiver operating characteristic curves in the validation datasets for the decision tree and smoothing splines analyses were 0.744, 0.713, 0.743, and 0.750 for model 1 and were 0.802, 0.714, 0.731 and 0.756 for model 2, for CART^®, MARS^®, RF™ and TREENET™, respectively.

Conclusion

The present study stands as a proof of concept for the use of classification tree algorithms for mortality prediction in an ICU setting.