MartÃn,J.D.(1), Soria,E.(1), Camps,G.(1), Serrano,A.J.(1), Pérez,J.J.(2), Jiménez,N.V.(2).
C/ Dr. Moliner, 50. 46100 Burjassot (València) - SPAIN
Introduction. Secondary anemia associated with the chronic renal failure (CRF) is a common clinical situation in patients who are receiving periodical hemodialysis. External administration of erythropoietin (rHu-EPO) is considered the best treatment for this pathology. Moreover, the pharmacokinetic and pharmacodynamic interindividual variability justifies the need for dosage individualisation.
Methods. We included 110 anemic patients with CRF in periodical hemodialysis with two or more EPO administrations per week and ferrum dosages less then 650 mg per month. They accepted taking part in the study. We excluded those patients whose etiology was the presence of multicystic kidney. At the end, the population is constituted by 77 patients (495 patterns) for obtaining the model and 33 more (174 patterns) for its validation.
We analyzed the influence of age, weight, EPO dosages, number of administrations per week, isoform of EPO (alpha or beta), ferritin, level of hemoglobin and ferrum dosage. We used time series methodology to predict the level of hemoglobin of the following month using the data of the previous two months.
Predictions are carried out with a multilayer feedforward neural network which constitutes a general non-linear regression model. These models are composed by a layered arrangement of artificial neurons in which each neuron of a given layer feeds all the neurons of the next layer. In addition, we have also used an Elman’s recurrent network in which hidden neurons (HN) are also connected with a context layer. Connections are adjusted using the backpropagation algorithm.
Results. The best results were obtained with a Multilayer Perceptron with 8 HN and an Elman’s network with 3 HN. Both models yielded RMS errors close to 0.2 g/dl, both for the training and the validation set. If we consider a correct prediction when the prediction error is smaller than the clinically relevant difference (0.25 g/dl), the percentage of success is higher than 97 % in both sets and with both models.
A user-friendly software was developed for clinical-decision aid.
Reference: PAGE 10 (2001) Abstr 214 [www.page-meeting.org/?abstract=214]
Poster: poster