Comparative study of NONMEM and Neural Networks for Cyclosporine Dosage Prediction in Renal Allograft Recipients

Camps-i-Valls, G.1; Porta-Oltra, B.2; Pérez-Ruixo, J. J.2; Soria-Olivas, E.1; Martín-Guerrero, J. D.1; Jiménez-Torres, N. V.2,3

Grup de Processat Digital de Senyals. Dept. Enginyeria Electrònica. Universitat de València. C/ Doctor Moliner, 50. 46100 Burjassot (València). Spain.

Purpose: The objective of the present study was to build accurate models in order to determine the next dosage of the cyclosporine A (CyA) in patients who have undergone kidney transplantation.

Patients: Thirty-six renal allograft patients were included in this study. Patients received a standard immunosuppressive regimen with a microemulsion lipidic formulation of CyA, mycophenolate mofetil (2g/d) and prednisone (0,5-1mg/kg per day). Patients were excluded if they received metabolic inducers or metabolic inhibitors, because they modify the pharmacokinetic profile of CyA. The initial oral dose of CyA (10 mg/Kg per day) was reduced according to the measured CyA blood concentration and the desired target range. Steady state CyA blood samples were withdrawn 12-14 hours after dose administration. CyA blood levels were measured by a specific monoclonal fluorescence polarization immunoassay (Abbott, TDx), with inter- and intra-assay variation coefficient of < 7,5%. We collected for each patient the following parameters: trough blood concentration (C), daily dose (DD), postoperative days, and several anthropometric, clinical and biochemical factors.

Methods: In order to quantify the relationship between DD and C empirical models based on non-linear mixed effects (NLME) modeling and artificial neural networks (ANN) were used. A time series methodology to predict the next dose using the previous data was employed. Mean prediction error (ME) was used as a measure of bias and the root mean squared error as a measure of precision.

Results: Results are shown in the table below.

Model

r

ME (CI 95%)

RMSE (CI 95%)

ANN

0.9852

0.0051

(0.0042, 0.0059)

0.3460

(0.3452, 0.3468)

A priori NONMEM prediction

0.9650

-0.0118

(-0.0554, 0.0319)

0.6154

(0.5029, 0.7102)

A posteriori NONMEM prediction

0.9840

-0.0120

(-0.0402, 0.0161)

0.3969

(0.2650, 0.4949)

Conclusions: It is concluded that neural networks are well-suited techniques for dosage prediction of CyA. A posteriori NONMEM predictions were similar in bias and precision although a priori NONMEM predictions were less accurate.

Reference: PAGE 10 () Abstr 217 [www.page-meeting.org/?abstract=217]

Poster: poster