Statistical approach to neural network model building for gentamicin peak predictions.

Feed forward neural networks are flexible, nonlinear modeling tools that are an extension of traditional statistical techniques. The hypothesis that feed forward neural network models can be built in a similar fashion as a statistical model was tested. Feed forward neural network models were built using forward and backward variable selection, and zero to five hidden nodes, and tanh and linear transfer functions were used. Gentamicin serum concentrations were predicted as a model drug for testing these methods. Peak observations from 392 patients were used to train, test, and validate the feed forward neural network. Inputs were demographic and drug dosing information. Model selection was performed using the Akaike information criteria (AIC), Bayesian information criteria (BIC), and a method of stopped training. The models with lowest root mean square (rms) error were those with all 10 inputs and five hidden nodes. Average rms error in the validation set was lowest for stopped training (1.46), then AIC (1.51), and finally BIC (1.56). Larger models tended to result in the best predictions. Overfitting can occur in models that are too large, either by using too many nodes in the hidden layer (rms = 1.49) or by using too many inputs with little information associated with them (rms = 1.70). We conclude that neural networks can be built using a large number of parameters that have good predictive performance. Care must be used during training to avoid overfitting the data. A stopped training method resulted in the network with the lowest rms error.