- Poster presentation
- Open Access
Prediction of hospital mortality by support vector machine versus logistic regression in patients with a haematological malignancy admitted to the ICU
© BioMed Central Ltd 2008
- Published: 13 March 2008
- Support Vector Machine
- Receiver Operating Characteristic
- Receiver Operating Characteristic Curve
- Haematological Malignancy
- Hospital Mortality
Prognostic indicators have been identified in critically ill patients with haematological malignancies [1, 2]. Locally developed risk-prediction models have proven to be equally or more accurate in predicting hospital mortality for these patients than models constructed for the general ICU population such as the APACHE II or SAPS II [1, 2]. The objective of this study is to compare the accuracy of predicting hospital mortality in patients with haematological malignancies admitted to the ICU between a risk-prediction model based on multiple logistic regression (MLR) and a support vector machine (SVM)-based risk prediction model.
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 and 100 were used for validation. In a first model, 12 input parameters were included for comparison between MLR and SVM. In a second, more complex model, 17 input parameters were used. MLR and SVM analyses were performed independently from each other. Discrimination was evaluated using the area under the receiver operating characteristic (ROC) curves (± SE).
The area under the ROC curve for the MLR and SVM in the validation set were 0.836 (± 0.04) versus 0.802 (± 0.04) in the first model (P = 0.21) and 0.891 (± 0.03) versus 0.808 (± 0.04) in the second, more complex model (P = 0.01), respectively. However, SVM only needed four variables to make its prediction in both models, whereas MLR needed seven and eight variables, respectively, in the first and second models.
MLR had better discriminative power for prediction of hospital mortality in critically ill patients with haematological malignancy as compared with SVM, but to the expense of inclusion of more input variables. The discriminative power in both models was sufficient for clinical use. After further validation and optimization, the application of SVM algorithms might contribute in the near future to the development of new risk-prediction models in the ICU.