PURPOSE: The purpose of this work was to predict plasma peak and trough levels of an aminoglycoside antibiotic in patients with severe illness in an intensive care unit by a novel approach. Plasma levels were predicted based on the values of 15 physiological measurements using an artificial neural network (ANN) simulator. METHOD: A data set of 15 physiological measurements for 30 patients was used to develop the model. The ANN structure consisted of three layers: an input layer comprised of 15 processing elements, a hidden layer comprised of 10 processing elements with a sigmoid function as an activation function, and an output layer of two processing elements (peak and trough levels). The weight between neurons was trained according to the delta rule back-propagation of errors algorithm. Predicted values were obtained by "leave-one-out" experiments by both ANN and multiple linear regression analysis (MLRA). RESULTS: The correlation coefficients between observed and predicted values obtained by ANN prediction using standardized data sets were r=0.825 and r=0.854 for peak and trough levels, respectively. The correlation coefficients obtained by MLRA were r=0. 037 and r=0.276 for peak and trough levels, respectively. These results indicate that ANN shows better performance in prediction of aminoglycoside plasma levels from patients' physiological measurements than MLRA. CONCLUSIONS: Prediction of plasma levels of antibiotic in patients with severe illness by ANN was superior to the standard statistical method. Standardization of input data was found to be important for better prediction. ANN has some advantages over standard statistical methods, as it can recognize complex relationships in the data.
PURPOSE: The purpose of this work was to predict plasma peak and trough levels of an aminoglycoside antibiotic in patients with severe illness in an intensive care unit by a novel approach. Plasma levels were predicted based on the values of 15 physiological measurements using an artificial neural network (ANN) simulator. METHOD: A data set of 15 physiological measurements for 30 patients was used to develop the model. The ANN structure consisted of three layers: an input layer comprised of 15 processing elements, a hidden layer comprised of 10 processing elements with a sigmoid function as an activation function, and an output layer of two processing elements (peak and trough levels). The weight between neurons was trained according to the delta rule back-propagation of errors algorithm. Predicted values were obtained by "leave-one-out" experiments by both ANN and multiple linear regression analysis (MLRA). RESULTS: The correlation coefficients between observed and predicted values obtained by ANN prediction using standardized data sets were r=0.825 and r=0.854 for peak and trough levels, respectively. The correlation coefficients obtained by MLRA were r=0. 037 and r=0.276 for peak and trough levels, respectively. These results indicate that ANN shows better performance in prediction of aminoglycoside plasma levels from patients' physiological measurements than MLRA. CONCLUSIONS: Prediction of plasma levels of antibiotic in patients with severe illness by ANN was superior to the standard statistical method. Standardization of input data was found to be important for better prediction. ANN has some advantages over standard statistical methods, as it can recognize complex relationships in the data.
Authors: S H Kang; M R Poynton; K M Kim; H Lee; D H Kim; S H Lee; K S Bae; O Linares; S E Kern; G J Noh Journal: Br J Clin Pharmacol Date: 2007-02-23 Impact factor: 4.335