An Example Of A Neural Network Like A Pharmacokinetic/Pharmacodynamic Model.

Serrano,A.J.(1), Martín,J.D.(1), Pérez,J.J (2), Camps,G.(1), Soria,E.(1), Jiménez,N.V.(2).

CL Dr. Moliner, 50. 46100 Burjassot (València) - SPAIN

Introduction. This work presents a novel approach to obtain pharmacokinetic/pharmacodynamic (PK/PD) model. Neural networks are non lineal models, and present a high flexibility to obtain a mapping between two sets. Another advantage is the absence of knowledge a priori to have a good performance. Moreover, several mathematical theorems proving that neural networks are universal approximators [1]. These characteristics show their great utility in PK/PD problems, also recent advances in these systems improve the performance of basic neural networks.

Objective. The emulative ability of neural network is evaluated with simulated low molecular weight heparin(LMWH) PK/PD model [2].

Methods. We simulated 4 different groups with 250 patients and 5 samples/patient (a: D 100 UI/12h; D 200 UI/24h; B 100 UI/12h and C 200 UI/24h). Two hundred patients per group were used to obtain the model and the rest to check. A multilayer perceptron with backpropagation algorithm was used. For PK prediction we used as inputs dose, interdose interval, time, baseline APTT and APTT. For PD prediction we used as inputs dose, interdose interval, time, baseline APTT and Anti-Xa. Mean of prediction error (ME) and root mean square prediction error (RMSE) was used to evaluate the bias and precision.

Results. The best models have 10 and 9 hidden neurons to the PD and PK problem, respectively. Correlation coefficients, between desired signal and output neural networks, were 0.96 (training and validation sets). In the training set of PK problem, ME is -0.0012 and RMSE is 0.1045 UI/mL. In the validation set of PK problem, ME is –0.0002 and RMSE is 0.1140 UI/mL. Results in PD problem are very similar.

References.
[1] Haykin, S. Neural Networks: A Comprehensive Foundation. Prentice-Hall. 1999.
[2] Schoemaker R. et al. Estimating impossible curves using NONMEM. Br J Clin Pharm 1996;42:283-90.

Reference: PAGE 10 (2001) Abstr 216 [www.page-meeting.org/?abstract=216]

Poster: poster