Prediction of the tacrolimus blood concentration in liver transplantation patients with support vector regression during an intensive care unit stay

Introduction

The tacrolimus blood concentration has wide intra-individual and inter-individual variability, especially in the initial phase after transplantation in the ICU. To insure clinical effect and to avoid side-effects, it is crucial to monitor concentrations very carefully. Prediction models can save time and resources, enabling clinicians and nurses to improve clinical care. The performance of linear and nonlinear support vector machines (SVM) as prediction models for the tacrolimus blood concentration in liver transplantation patients is compared with linear regression analysis.

Methods

Five hundred and twenty-three tacrolimus blood concentration levels, together with 35 other relevant variables from 56 liver transplantation patients between 2002 and 2006, were extracted from Ghent University Hospital database (ICU Information System IZIS) (Centricity Critical Care Clinisoft; GE Healthcare). Multiple linear regression, and support vector regression with linear and nonlinear (RBF) kernel functions were performed, after selection of relevant data components and model parameters. Performances of the prediction models on unseen datasets were analyzed with fivefold cross-validation. Wilcoxon signed-rank analysis was performed to examine differences in performances between prediction models and to analyze differences between real and predicted tacrolimus blood concentrations.

Results

The mean absolute difference with the measured tacrolimus blood concentration in the predicted regression model was 2.34 ng/ml (SD 2.51). Linear SVM and RBF SVM prediction models had mean absolute differences with the measured tacrolimus blood concentration of, respectively, 2.20 ng/ml (SD 2.55) and 2.07 ng/ml (SD 2.16). These differences were within an acceptable clinical range. Statistical analysis demonstrated significant better performance of linear (P < 0.001) and nonlinear (P = 0.002) SVM (Figure 1) in comparison with linear regression. Moreover, the nonlinear RBF SVM required only seven data components to perform this prediction, compared with 10 and 12 components needed, respectively, by multiple linear regression and linear SVM.

Figure 1

Conclusion

Performance of SVM with linear and nonlinear kernel function was excellent and superior in comparison with the multiple linear regression model in predicting the tacrolimus blood concentration.