Home About us Current issue Back issues Submission Instructions Advertise Contact Login   

Search Article 
Advanced search 
Saudi Journal of Kidney Diseases and Transplantation
Users online: 2180 Home Bookmark this page Print this page Email this page Small font sizeDefault font size Increase font size 

Table of Contents   
Year : 2011  |  Volume : 22  |  Issue : 4  |  Page : 705-711
Artificial neural network for prediction of equilibrated dialysis dose without intradialytic sample

1 Electrical Communication & Electronics Systems Engineering department, Modern Science and Arts University (MSA), 6th of October City, Egypt
2 Systems and Biomedical Engineering Department, Faculty of Engineering, Cairo University, Egypt

Click here for correspondence address and email

Date of Web Publication9-Jul-2011


Post-dialysis urea rebound (PDUR) is a cause of Kt/V overestimation when it is calculated from pre-dialysis and the immediate post-dialysis blood urea collections. Measuring PDUR requires a 30-or 60-min post-dialysis sampling, which is inconvenient. In this study, a supervised neural network was proposed to predict the equilibrated urea (C eq) at 60 min after the end of hemodialysis (HD). Data of 150 patients from a dialysis unit were analyzed. C eq was measured 60 min after each HD session to calculate PDUR, equilibrated urea reduction rate eq (URR), and ( eq Kt/V). The mean percentage of true urea rebound measured after 60 min of HD session was 19.6 ± 10.7. The mean urea rebound observed from the artificial neural network (ANN) was 18.6 ± 13.9%, while the means were 24.8 ± 14.1% and 21.3 ± 3.49% using Smye and Daugirdas methods, respectively. The ANN model achieved a correlation coefficient of 0.97 (P <0.0001), while the Smye and Daugirdas methods yielded R = 0.81 and 0.93, respectively (P <0.0001); the errors of the Smye method were larger than those of the other methods and resulted in a considerable bias in all cases, while the predictive accuracy for ( eq Kt/V) 60 was equally good by the Daugirdas' formula and the ANN . We conclude that the use of the ANN urea estimation yields accurate results when used to calculate ( eq Kt/V).

How to cite this article:
Azar AT, Wahba KM. Artificial neural network for prediction of equilibrated dialysis dose without intradialytic sample. Saudi J Kidney Dis Transpl 2011;22:705-11

How to cite this URL:
Azar AT, Wahba KM. Artificial neural network for prediction of equilibrated dialysis dose without intradialytic sample. Saudi J Kidney Dis Transpl [serial online] 2011 [cited 2019 Jul 22];22:705-11. Available from: http://www.sjkdt.org/text.asp?2011/22/4/705/82647

   Introduction Top

Post-dialysis urea rebound (PDUR) is due to the inter-compartmental equilibration of urea during hemodialysis (HD) and supports the concept of an intradialytic two-pool model to calculate the dialysis dose. These compartments are considered to be extracellular and intra-cellular. [1],[2] PDUR is a critical problem, which influences the calculation of equilibrated urea (C eq) and increases the risk of inaccurate estimation of the dose of dialysis; [3] Kt/V is greater than that actually achieved in the HD patients when calculated using the immediate post-dialysis blood urea concentration. [4] Because a delay of 30-60 min after dialysis before sampling the urea is inconvenient for both the clinician and patients, several methods have been devised to predict the PDUR-equilibrated Kt/V. [5],[6],[7],[8],[9],[10],[11],[12],[13]

Smye et al [5],[13] predicted C eq concentration using the principle of mass conservation and then employed this C eq to calculate the equilibrated dialysis dose ( eq Kt/V). Daugirdas [6],[7],[8] proposed a simpler method, where a single-pool Kt/V is modified according to the speed of dialysis (K/V) to obtain a double-pool estimate of Kt/V. These two approaches make some biological assumptions that are not always true, but remain the most accepted models.

Recently, Guh et al applied an artificial neural network (ANN) method to deal with this problem using the AINET neural network, which is a self-organizing system derived from a probabilistic approach. [14] In this method, a supervised neural network model was proposed to predict the C eq 60 min after the end of HD and we also employed the Bland-Altman agreement method to compare the different predictive methods. [15] The ANN is the well-known multilayer perceptron (MP) trained with the Levenberg-Marquardt (LM) algorithm as shown in [Figure 1]. [16],[17],[18]
Figure 1: The artificial neural network (ANN) model.

Click here to view

We aim in this study to evaluate the ANN utilization in predicting the rebound of urea and the different input parameter combinations in order to find the most predictive ones and all the results are compared with those of the Smye and Daugirdas models.

   Methodology Top

We studied 156 stable HD patients in a dialysis unit. The selection criteria included: (a) patients without infections or hospitalizations in the previous 30 days and (b) all the patients who had A-V fistulas with blood flow rate (Q B ) ≥300 mL/min. All the patients received HD three times a week with Fresenius machines (model 4008B; Fresenius Medical Care, Bad Homburg, Germany) with variable bicarbonate and sodium. Hollow-fiber polysulfone (1.3 m 2 and 1.6 m 2 low flux dialyzers, Fresenius Medical Care, Bad Homburg, Germany) dialyzers were used. For the purpose of this study, all patients were dialyzed over 240 min, and blood and dialysate flows were fixed at 300 and 500 mL/min, respectively. The mean age of the patients was 48.2 ± 13.4 years, and the mean pre-dialysis body weight ( pre BW) was 68.8 ± 15.5 kg. The mean ultrafiltration (UF) (difference between pre-and post-dialysis body weights) was 8.54 ± 5.15 mL/min. All blood samples were obtained at the midweek HD session. Four blood samples were obtained from the arterial line at different times for urea determinations: (1) urea pre-dialysis (C pre ) at the beginning of the procedure, (2) intradialysis urea (C int ) at the middle of the HD session, (3) urea post-dialysis (C post ) at the end of the HD session, and (4) equilibrated urea (C eq ) at 60 min after the end of HD. To extract the blood samples for the determination of the intradialysis urea (C int ) and post-dialysis urea (C post ), the Q B was slowed to 50 mL/min and the blood was extracted 15 s later. In this situation, access recirculation stops or ceases, and the dialyzer inlet blood reflects the arterial urea concentration. The intradialysis sample (C int ) was obtained 80 min after the beginning of the HD session in order to make the results comparable with those reported by Guh et al [14] and because Smye et al [5] originally proposed that the intradialysis sample be obtained between 60 min after the beginning of the session and before the last 20 min of the end of it.

A. Methods for estimating equilibrated urea concentration (C eq )

1. Gold Standard method: The true C eq measured from the blood samples taken at 60 min after the end of the HD session was used as a "gold standard" for post-dialysis C eq and urea rebound (UR) was calculated as:

2. Smye method: [5] The value of equilibrated BUN (C eq ) was calculated from the Smye method according to the following equation:

where C eq is the concentration at the equilibrium; C o, C s , and C post are urea concentrations at the start, at time s and at the end of dialysis, respectively; t is the dialysis time (in minutes), t s is the time when the C s sample was taken.

3. ANN Method: ANN is proposed in this paper as a predictor of the C eq . NNs are mathematical models that grew out of early attempts to model the behavior of the human central nervous system. They usually consist of a set of simple processing units ("neurones") that are highly interconnected through coefficients called weights ("synapses"). The NNs resemble the brain in two main aspects: (a) knowledge is acquired by the network from its environment through a learning process and (b) interneuron connection strengths, known as synaptic weights, are modified during the learning process and used to store the acquired knowledge. [17] The chosen model was the MP, which is a feedforward ANN model with one input layer (equal in size to the number of input variables), hidden layers (none, one, or several), and an output layer. [17],[18] Each layer is connected by numerical coefficients called weights resembling neural synapses. The ANN was trained with a modified back propagation algorithm, which is a supervised learning strategy that consists of the modification of the weight values in order to minimize an error function, when an input/ output pair is presented to the net. [17],[19] The LM training algorithm was used, which is an intermediate optimization algorithm between the Gauss-Newton method and gradient descent algorithm, addressing the shortcomings of each of those techniques. [16] LM uses gradient descent to improve on an initial guess for its parameters and transforms to the Gauss-Newton method as it approaches the minimum value of the cost function. Once it approaches the minimum, it transforms back to the gradient descent algorithm to improve the accuracy. The net was trained to predict the C eq from the following independent variables: Pre-dialysis BUN (C pre ), dialysis time in minutes (T d ), blood flow rate (BFR), ultrafiltration rate (UFR) and post-dialysis BUN (C post ) as shown in [Figure 2]. Different net sizes were tested in order to look for the most appropriate number of hidden nodes. Once the net size selection was made, different parameter combinations were tested to build the ANN model, in order to see their influence in the prediction process and evaluate their importance in the predictive model. The hidden layers were given a logistic sigmoid transfer function and the output layer consisted of a single node with a linear transfer function.
Figure 2: Comparison plot between equilibrated-BUN and other prediction methods.

Click here to view

B. Equilibrated Kt/V formulas

A single-pool Daugirdas formula was used to calculate the single-pool Kt/V (Kt/V sp ): [7]

where R = U post /U pre , UF is ultrafiltration, W is the post-dialysis body weight, and t is the duration of the dialysis session in hours.

1. Gold standard Method: To derive the "gold standard" equilibrated Kt/V ( eq Kt/V), we use the Kt/V sp of equation 3, substituting the "equilibrated" postdialysis BUN for the immediate post-BUN when computing R:

where the C eq60 is the measured urea concentration from a blood sample taken at 60 min after the end of HD. The eq (Kt/V) 60 was used as a reference for comparison purposes.

2. Smye Method: The equilibrated Kt/V using Smye method can be calculated by the following formula:

where C eqSmye is the estimated C eq with the intradialysis urea sample explained above.

3. The equilibrated Kt/V using ANN method can be calculated by the following formula:

where the C eqANN is the ANN estimation of C eq .

4. The double-pool Daugirdas formula [7] was also used to estimate the eq Kt/V ( eq Kt/V Dau ):

All the estimated eq Kt/V ( eq Kt/V Dau , eq Kt/V Smye and eq Kt/V ANN ) were compared against the "gold standard equilibrated Kt/V" eq (Kt/V) 60.

C. Training process of the ANN

In the net size selection process, complete data were randomly mixed and two training/ test set pairs were built as follows: (A) training: first 70% of the samples, test: last 30% of the samples; (B) training: last 70% of the samples, test: first 30% of the samples. In this way, the jackknife method [18] is applied for validation purposes. The mean squared error (MSE) between the net's output and the C eq of the test sets was averaged over ten training procedures (five for each training/test set pair) and the net size, which achieved the lowest MSE, was selected. Once the net size and the input parameters were selected, a new net was trained with an early stopped method to improve generalization. In this technique, the available data were divided into three subsets. The first subset is the training set, which is used for computing the gradient and the network weights (60% of the samples). The second subset is the validation set (20% of the samples). The error on the validation set is monitored during the training process. The third subset is the test set (20% of the samples). The validation error normally decreases during the initial phase of training, as does the training set error. However, when the network begins to over fit the data, the error in the validation set typically begins to rise. When the validation error increases for a specified number of iterations, the training is stopped, and the weights and biases at the minimum of the validation error are returned. [16] This net (ANN) was used to predict the C eq .

D. Comparison analysis

All the methods (the selected ANN, Smye and Daugirdas methods) were compared against the C eq measured from the blood samples taken at 60 min after the end of the HD session. In all cases, the following coefficients were used: (i) correlation: index of direct association between two methods and (ii) Bland-Altman agreement method: [8] this method is based on the analysis of the average and the standard deviation of the difference between the model and the "gold standard", against the true value ("gold standard"). All methods of equilibrated Kt/V were compared using the same analysis as explained above. The results were compared against the "gold standard Kt/V" eq (Kt/V) 60 equation 4.

   Results Top

An ANN with two hidden nodes was selected because it achieves the lowest MSE and is robust to different initialization procedures of the ANN initial weights.

The ANN model showed more accuracy and better correlation coefficient (R) than the Smye method [Table 1]. In all the patients, the ANN achieved an R = 0.91 versus an R = 0.75 (P <0.0001) obtained with the Smye method.
Table 1: Association between true equilibrated urea and other methods.

Click here to view

The results of urea rebound prediction methods are summarized in [Table 2]. The mean percentage of true urea rebound measured after 60 min of the HD session was 19.6 ± 10.7. The mean urea rebound observed from the ANN was 18.6 ± 13.9%, while the means of urea rebound using Smye and Daugirdas methods were 24.8 ± 14.1% and 21.3 ± 3.49%, respectively. The ANN model achieved an R = 0.97, while the Smye and Daugirdas methods yielded R = 0.81 and 0.93, respectively (P <0.0001). The percentage errors of the Smye method were bigger than the percentage errors achieved by Daugirdas and ANN methods for both the mean and the standard deviation.
Table 2: Association between true urea rebound and other methods.

Click here to view

The C eq estimated from the Smye method and ANN model was used to estimate the eq (Kt/V). eq (Kt/V) was also estimated with double-pool Daugirdas formula (equations 7 and 8). All these estimates were compared against the "gold standard KtV" eq (Kt/V) 60 calculated using equation 4. The true eq (Kt/V) 60 was 1.13 ± 0.34, while the the eq (Kt/V) using Smye, Dauigirdas and ANN were 0.93 ± 0.46, 1.08 ± 0.54 and 1.17 ± 0.23, respectively. [Table 3] shows that the value of the calculated eq (Kt/V) using the C eq Smye estimation was the smallest of those calculated by the other methods. In addition, Smye had the smallest correlation coefficient and the biggest percentage error, not only showing a considerable bias in all cases, but also showing the highest level of inaccuracy. Finally, the predictive accuracy for eq (Kt/V) 60 was equally good by the Daugirdas and the ANN formulas.
Table 3: Association between true equilibrated Kt/V and other methods.

Click here to view

   Discussion Top

ANN has been successfully applied in many biomedical problems and we have demonstrated in this study their ability to accurately predict the C eq with the corresponding benefits for the true dose calculations for HD patient without the need for an intradialytic sample. The ANN model is more accurate than the Smye formula because the ANN model showed, from a Bland-Altman point of view, lower mean and standard deviation errors. The analysis of % error against the % of urea rebound showed that the ANN is very robust to the level of urea rebound. The Smye method overestimated for % UR by about 26.6%. The eq Kt/V estimated with the C eq calculated with the Smye formula is not appropriate because it shows a great dispersion. The Bland-Altman analysis showed that the Smye method underestimated the eq (Kt/V) 60 by about 17.7%. This bad result could be explained by the consideration discussed above. The Daugirdas rate equation is accurate and the results agree with the results of the HEMO study. [12] The eq (Kt/V) calculated with the C eq estimated by the ANN is also accurate, and it showed a lower bias and greater robustness to noisy data. However, the ANN model performed better than the Daugirdas method as shown in [Figure 3]. It was found that the rate equation only slightly underestimated true eq (Kt/V) 60 by about 4.8%, while the ANN method overestimated the true values by about 3.7% as shown in [Figure 4].
Figure 3: Comparison Plot between True PDUR and other methods.

Click here to view
Figure 4: Comparison plot between true equilibrated Kt/V and other methods.

Click here to view

The intrinsic nonlinearities involved in the ANN model render this approach very promising to directly estimate the eq (Kt/V) without the previous estimation of the C eq . The results presented here suggest that the ANN, based on limited clinical parameters, is an excellent alternative method to other traditional urea kinetic models (UKM) for accurately predicting equilibrated urea concentrations and urea rebound in HD patients. Future studies should focus on the evaluation and analysis of new noninvasive variables in order to find the most significant ones. We conclude that this study found the ANN model a very promising analysis tool to estimate urea rebound in HD patients.

   Acknowledgments Top

The authors thank all the medical staff at the Nephrology Department in Ahmad Maher Teaching Hospital, Cairo, Egypt, for their invaluable support during the course of this study.

   References Top

1.Popovich RP, Hlavinka DJ, Bomar JB, Moncrief JW, Decherd JF. The consequences of physiological resistance on metabolite removal from the patient-artificial kidney system. Trans Am Soc Artif Intern Organs 1975;21:108-15.  Back to cited text no. 1
2.Sargent JA, Gotch FA. Principles and bio-physics of dialysis, in Replacement of Renal Function by Dialysis (4 th ed), edited by Jacobs C, Kjellstrand CM, Koch KM, Winchester JF, Dordrecht, Kiuwer Academic, 1996. p. 34-102.  Back to cited text no. 2
3.Alloati S, Molino A, Manes M, Bosticardo GM. Urea rebound and effectively delivered dialysis dose. Nephrol Dial Transplant 1998;13 (suppl 6):25-30.  Back to cited text no. 3
4.Gotch F, Sargent A. A mechanistic analysis of the National Cooperative Dialysis Study. Kidney Int 1985;28:526-34.  Back to cited text no. 4
5.Smye SW, Dunderdale E, Brownridgr G, Will E. Estimation of treatment dose in high-efficiency hemodialysis. Nephron 1994;27:24-9.  Back to cited text no. 5
6.Daugirdas JT, Schneditz D. Overestimation of hemodialysis dose dependent on dialysis efficiency by regional blood flow but not by conventional two pool urea kinetic analysis. ASAIO J 1995;41:M719-24.  Back to cited text no. 6
7.Daugirdas JT. Simplified equations for monitoring Kt/V, PCRn, eKt/V, and ePCRn. Adv Ren Replace Ther 1995;2:295-304.  Back to cited text no. 7
8.Daugirdas JT, Burke MS, Balter P, Priester-Coary A, Majka T. Screening for extreme post-dialysis urea rebound using the Smye method: Patients with access recirculation identified when a slow flow method is not used to draw the post-dialysis blood. Am J Kidney Dis 1996;28:727-31.  Back to cited text no. 8
9.Tattersall JA, Detakats D, Chamney P, Greenwood RN, Farrington K. The post-hemodialysis rebound: Predicting and quantifying its effect on Kt/V. Kidney Int 1996;50:2094-102.  Back to cited text no. 9
10.Maduell F, Garcia-Valdecasas J, Garcia H, et al. Validation of different methods to calculate Kt/V considering post-dialysis rebound. Nephrol Dial Transplant 1997;12:1928-33.  Back to cited text no. 10
11.Canaud B, Bosc JY, Leblanc M, et al. A simple and accurate method to determine equilibrated post-dialysis urea concentration. Kidney Int 1997;51:2000-5.  Back to cited text no. 11
12.Daugirdas JT, Depner TA, Gotch FA, et al. Comparison of methods to predict equilibrated Kt/V in the HEMO Pilot Study. Kidney Int 1997;52:1395-404.  Back to cited text no. 12
13.Kaufman AM, Schneditz D, Smye SW, Polaschegg HD, Levin NW. Solute disequlibrium and multi-compartment modeling. Adv Ren Replac Ther 1995;2:319-29.  Back to cited text no. 13
14.Guh J, Yang C, Yang J, Chen L, Lai Y. Prediction of equilibrated postdialysis BUN by an artificial neural network in high-efficiency hemodialysis. Am J Kidney Dis 1998;31:638-46.  Back to cited text no. 14
15.Bland JM, Altman DG. Statistical methods for assessing agreement between two methods of clinical measurement. Lancet 1986;8:307-10.  Back to cited text no. 15
16.Hagan M, Menhaj M. Training feed-forward networks with the Marquardt algorithm. IEEE Trans Neural Netw 1994;5:989-93.  Back to cited text no. 16
17.Haykin S. (Ed.) Neural networks. A comprehensive foundation, 2 nd ed. (Prentice Hall, New Jersey, 1999.  Back to cited text no. 17
18.Cross SS, Harrison RF, Kennedy RI. Introduction to neural networks. Lancet 1995;346:1075-9.  Back to cited text no. 18
19.Erb RJ. Introduction to back propagation neural network computation. Pharm Res 1993;10:165-70.  Back to cited text no. 19

Correspondence Address:
Ahmad Taher Azar
Menoufeya, Menouf, Ahmad Orabi Square, Azar Building
Login to access the Email id

PMID: 21743214

Rights and Permissions


  [Figure 1], [Figure 2], [Figure 3], [Figure 4]

  [Table 1], [Table 2], [Table 3]

This article has been cited by
1 Adaptive network based on fuzzy inference system for equilibrated urea concentration prediction
Azar, A.T.
Computer Methods and Programs in Biomedicine. 2013; 111(3): 578-591
2 Neuro-fuzzy applications in dialysis systems
Azar, A.T.
Studies in Computational Intelligence. 2013; 405: 1223-1274
3 Artificial neural networks applications in dialysis
Fernández, E.A. and Valtuille, R. and Balzarini, M.
Studies in Computational Intelligence. 2013; 405: 1145-1179
4 Double pool urea kinetic modeling
Azar, A.T. and Yashiro, M. and Schneditz, D. and Roa, L.M.
Studies in Computational Intelligence. 2013; 404: 627-687


    Similar in PUBMED
    Search Pubmed for
    Search in Google Scholar for
    Email Alert *
    Add to My List *
* Registration required (free)  

    Article Figures
    Article Tables

 Article Access Statistics
    PDF Downloaded341    
    Comments [Add]    
    Cited by others 4    

Recommend this journal