Clinical Application of Artificial Neural Network to Predict Creatinine Clearance in Elderly Inpatients

(1)

Clinical Application of Artificial Neural

Network to Predict Creatinine Clearance in

Elderly Inpatients

Chiou-An Chena,d, Tsung-Ming Hub, Yen-Yu Chenc, Yu-Chuan Lid, Jainn-Shiun Chiue,1

a

Department of Internal Medicine, Tri-Service General Hospital, Taipei, Taiwan b

Department of Adult Psychiatry, Yuli Hospital, Hualien, Taiwan c

Department of Neurology, Changhua Christian Hospital, Changhua, Taiwan d

Graduate Institute of Medical Informatics, Taipei Medical University, Taipei, Taiwan e

Department of Nuclear Medicine, Dalin Tzu Chi General Hospital, Chiayi, Taiwan

Abstract. Serum creatinine (Cr) is commonly used as an index of glomerular

filtration rate but easily inaccurate affected by aging or malnutrition. We therefore developed an artificial neural network (ANN) to predict 24-hour Cr clearance (CCr) in 22 elderly inpatients. Demographic records, anthropometric measurements, and blood biochemistries were collected as predictors. 24-hour CCr predicted by three conventional algebraic equations (Cockcroft-Gault formula, Jelliffe formula, Modification of Diet in Renal Disease formula) and ANN (CCr-ANN) were separately compared with measured 24-hour CCr (CCr-24H). CCr-24H (60.31 ± 10.65 ml/mim/1.73m2) and CCr-ANN (61.49 ± 10.20 ml/mim/1.73m2) was insignificantly different (p = 0.47). CCr-24H and CCr-ANN were statistically correlated (r = 0.94, p < 0.001) with less bias (peak centered most closely to zero with shortest tails in empirical cumulative distribution plot) and better goodness-of-fit (root mean square error = 16.01) than other equations. ANN is a promising application to accurately predict 24-hour CCr in elderly inpatients. Keywords: Neural network, Creatinine Clearance, Aging.

1. Introduction

The aging population has been growing rapidly corresponding with a simultaneous increase of hospitalization rate. Elders admitted to hospital often take many medications to manage acute illness. About two thirds of the elders can not maintain their normal renal function whilst most of them have only 50% of glomerular filtration rate (GRF) [1]. Renal function deterioration makes them vulnerable to prescribed agents, particularly many of which are renally excreted with nephrotoxicity. Therefore, accurate evaluation of renal function in elders with acute illness is essential for dosage adjustment and appropriate balance fluid and electrolytes.

Serum creatinine (Cr) is commonly used as a clinical index of GFR to present renal function but it is easily affected by aging or malnutrition. Although the reference method to determine GFR is measuring inulin or radiopharmaceuticals clearance, these

1_{Corresponding Author: Jainn-Shiun Chiu, MD. Department of Nuclear Medicine, Dalin Tzu Chi General}

(2)

techniques are impractical in most hospitals due to burdensome inconvenience to patients, time-consuming and high-cost procedures, requirement of complex equipments with specialized staffs, or radiation exposure. In clinical practice, 24-hour urinary collection for Cr clearance (CCr) is the most recognized method to measure renal function as a rational computation of GFR. On account of carelessness, forgetfulness, confusion or urinary incontinence, exact 24-hour urinary collection in elderly patients could be a tough mission. Consequently, several algebraic equations are used to predict CCr for dosing renally eliminated drugs. Even though algebraic equations are widely adopted as alternatives of renal function across all ages, they may fail to predict CCr in elderly patients. Hence, a novel method to predict 24-hour CCr would be very useful. With the advances of software engineering, it is easy to use artificial neural network (ANN) which is a computational simulation of biologic nervous system. The purpose of present study was to evaluate the ability of ANN in predicting 24-hour CCr for elderly inpatients and compared its predictive performance with three algebraic equations.

2. Material and Methods

2.1. Participants

Inpatients more than or equal to 60 years of age, admitted to acute ward, were enrolled in the study if they had urinary bladder catheters for at least 48 hours before enrollment. Serum Cr measured the week before, during, and after the collection indicated that all participants had stable renal function with variation less than 17.68 μmol/L. Inpatients who take drugs affecting serum Cr concentration were also excluded. The final study population had 22 elderly inpatients.

2.2. Measurements

An accurate 24-hour urinary collection was achieved and its volume was recorded, together with serum urea nitrogen, serum Cr, serum albumin, urinary Cr, and patient’s weight and height. Urine flow rate was calculated from total urine volume (ml) divided by collection period (1440 minutes). Blood collection was carried out in the morning, immediately following urine collection. As CCr is established upon an average body surface area (BSA) of 1.73 m2, we corrected the CCr values for each participant according to individual BSA. This was calculated using DuBois formula: BSA = weight0.425 × height0.725 × 0.007184, where weight is in kilogram and height is in centimeter [2]. Afterwards, 24-hour CCr was calculated as conventional: CCr-24H = (urinary Cr concentration × urinary volume × 1.73) / (serum Cr concentration × 1440 × BSA). Three algebraic equations used consisted of 24-hour CCr estimated as Cockcroft-Gault formula (CCr-CG) [3], Jelliffe formula (CCr-JF) [4], and Modification of Diet in Renal Disease (MDRD) formula (CCr-MDRD) [5].

2.3. ANN construction

We used STATISTICA 7 (StatSoft, Inc., Oklahoma, US) to create various formulations of ANN models. A built-in intelligent problem solver was utilized to choose the excellent ANN. Demographic records (gender, age), anthropometric measurements

(3)

(height, weight), and serum biochemistries (urea nitrogen, Cr, albumin) were employed as input variables into ANN models; measured CCr-24H was entered as output variable for supervised training algorithm. We performed the leave-one-out cross validation for resampling. This technique requires a substantial number of experiments, but it can compute an average score over different partitions with an accurate estimate of generalization performance which is destined to avoid possible bias introduced by depending on any one particular division into test and training parts.

During the training process, the automatic network designer activated a heuristic and optimized approach to determine the proper arrangement. It managed a large amount of tests, which were used to decide the best architecture. It could compare linear network model, multilayered perceptron model, radial basis function model, probabilistic and generalized regression neural network model. To compare the performance of networks with different input variables, the network designer balanced error against type and diversity as criteria to select networks.

Using sensitivity analysis, the intelligent problem solver calculated the predictive error ratio for each input variable according to their degree of validity [6]. The ratio was expressed as the relation between the predictive errors for the model with removed input variables and the total network errors calculated based on all input variables. This index represented that the contributions made by different input variables in predicting the outcome were ranked in order of descending importance. After the network was allowed to run, CCr-ANN was correlated with CCr-24H; and if the ANN predicted the outcome incorrectly, by a process of back propagation, hidden weights within the network were readjusted until the predicted outcome was accurate. At last, the intelligent problem solver retained the best architecture and the optimum set of input variables.

2.4. Statistical analysis

Data were expressed as mean ± standard error. Correlations between each input variable and measured 24-hour CCr were analyzed by Spearman's rank correlation coefficient (Rs). Predictive 24-hour CCr derived from algebraic equations and ANN were compared with measured 24-hour CCr by Wilcoxon test. To test the predictive performance, the association between measured and predicted 24-hour CCr was also expressed in terms of Rs. High correlation means the measurements by two methods are linearly related; however, high correlation does not mean the agreement. Folded empirical cumulative distribution plot is a useful graphical analysis to display lack of agreement by presenting bias as the peak (median) and the difference as two tails (lowest and highest values) [7]. If two assays are unbiased with respect to each other, the peak will be centered over zero. Long tails in the plot reflect large differences between the methods. Although this plot can be used as an index of bias, a poor goodness-of-fit can occur in a lesser bias model. To avoid this problem, we use root mean square error (RMSE) as a measure of goodness-of-fit. The model with smaller RMSE value has better fit if there is more than one model to fit the data. Significance was defined in our study as p value less than 0.05.

3. Results

The characteristics of enrolled patients are presented in Table 1. Among these input variables, urea nitrogen, Cr, and albumin were statistically correlated with CCr-24H. The final best ANN which is a multilayer perceptron network with one input layer of 3

(4)

neurons, first hidden layer of 5 neurons, second hidden layer of 1 neuron and one output layer with 1 neuron. Four variables (age, gender, height, weight) were pruned after training processes and then remaining 3 input variables (urea nitrogen, Cr, albumin) were adopted as significant predictors. In order of descending importance, urea nitrogen, Cr, and albumin made the most contributions which their ratios of sensitivity analysis in approved ANN were 2.73, 2.49, and 1.07.

Table 1. Characteristics of enrolled patients.

Characteristics Value or ratio Rs P value

Age (years) 80.05 ± 1.46 -0.01 0.80

Gender (male/female) 14/8 0.07 0.73

Height (cm) 163.45 ± 1.34 -0.04 0.84

Weight (kg) 60.60 ± 1.81 -0.07 0.76

Urea nitrogen (mmol/L) 18.82 ± 3.68 -0.88 < 0.001

Cr (μmol/L) 139.67 ± 27.40 -0.92 < 0.001

Albumin (g/L) 28.8 ± 0.91 0.53 0.02

Table 2. Results of 24-hour CCr by measurement, algebraic equations, and ANN.

24-hour CCr (ml/min/1.73m2) P value

CCr-24H 60.31 ± 10.65 -

CCr-CG 74.48 ± 14.08 0.08

CCr-JF 72.52 ± 13.86 0.11

CCr-MDRD 91.86 ± 19.30 0.01

CCr-ANN 61.49 ± 10.20 0.47

Table 3. Spearman's rank correlation coefficient, folded empirical cumulative distribution plot, and RMSE for various models compared with measured 24-hour CCr.

CCr-CG CCr-JF CCr-MDRD CCr-ANN

Rs 0.91 0.91 0.90 0.94

Folded empirical cumulative distribution plot

Median -7.86 -7.90 -11.98 -3.25

Lowest value -92.30 -108.94 -188.67 -36.28

Highest value 49.13 43.61 30.19 31.90

RMSE 37.54 36.53 61.71 16.01

Table 2 shows results of measured and predicted 24-hour CCr. Although the predicted value for all methods overestimated the measured 24-hour CCr, no statistical difference was found except MDRD formula. As shown in Table 3, all estimates significantly correlated with measured 24-hour CCr (p < 0.001). Nevertheless, ANN had the peak (-3.25) centered most closely to zero with shortest tails (-36.28 ~ 31.90) which represented smallest bias than other equations (Figure 1). The RMSE value of ANN was also smallest which represented that ANN had better fit than other equations (Table 3).

4. Discussion

Many physicians like to use serum Cr level as index of renal function in clinical practice, but there are several significant factors affecting serum Cr concentration. GFR has

(5)

declined to approximately half the normal level before the serum Cr concentration rises above the upper limit of normal. In the developed countries, the ever-expanding aged population with improvement in life expectance will both translate into further increase in the number of elderly inpatients. Some of them have chronic diseases and usually take several medications. Acute illness may further compromise their limited renal function. Since many physicians prescribe the drugs with empirical dosage, these elderly inpatients may receive inadequate or toxic dosage. Therefore, close monitoring of renal function is warranted. In view of daily practice, predictive algorithms were needed to estimate renal function. Herein, we constructed a novel approach using ANN to predict 24-hour CCr against measured values on a reliable 24-hour urine collection.

Figure 1. Folded empirical cumulative distribution plot between 24-hour CCr predicted by three algebraic equations and ANN on the basis of measured 24-hour CCr.

This ANN proved to have better performance in predicting 24-hour CCr than algebraic equations. All algebraic equations excessively overestimated 24-hour CCr even though only the estimate by MDRD formula was statistically higher. Nevertheless, the difference between 24-hour CCr predicted by ANN and direct measurement was statistically insignificant. Although all models correlated with measured 24-hour CCr, algebraic equations fitted to a relatively lower correlation (Rs = 0.90 to 0.91) than ANN (Rs = 0.94). This indicated that algebraic equations based on linear relationship are not the best models. Because nonlinearity is essential in medicine, ANN has such an advantage to predict complex nonlinear relationships in biological processes. Furthermore, the performance of ANN could ceaselessly become better over time because ANN can dynamically retrain and update their topology if more data is gathered.

For any predictive model to be feasible in making clinical decisions, the most importance is only data that are easily available to the physicians at the time of triage. Therefore, we selected clinical variables to be possible predictors. Through univariate analysis, three significant biochemical predictors (urea nitrogen, Cr, and albumin) were identified. Of note, ANN may recognize input variables that are most valuable with regard to accuracy of prediction. Our ANN finally retained the same biochemical predictors to generate better prediction than those by algebraic equations. This finding might support that biochemical predictors could truly reflect the underlying value of 24-hour CCr in elderly inpatient. The inference is compatible with the clinical conditions

(6)

in elderly inpatients with diverse patterns of blood urea nitrogen and Cr, or poor nutritional status. Commonly, it is difficult to measure weight and height in bed-ridden inpatients. Hence, this ANN can use biochemistries with ease to execute the prediction.

The studies of ANN to predict 24-hour CCr are limited in clinical medicine. In a similar study, Herman et al. developed an ANN with 16 inputs to predict 24-hour CCr in patients with human immunodeficiency virus and compared its predictability with algebraic equations [8]. The performance of their ANN was superior to that of algebraic equations, but our model had simple structure with 3 predictors to obtain a comparable accuracy. However, the different study population may result in this distinction. In another study, Gabutti et al. constructed an ANN with the inclusion of cystatin C in the Cockcroft-Gault formula to predict 24-hour CCr in inpatients with mildly impaired renal function. Their results showed a statistically significant reduction in the predictive error [9]. Nevertheless, this procedure probably has no clinical relevance since cystatin C is not clinically available in many hospitals or medical laboratories.

Although the numbers of participants were limited in our study, we utilized resampling method to overcome this difficulty and it showed the superior results. However, much larger numbers of elderly inpatients are still needed in order to create more reliable models. The current model will also require external validation in a new group of elderly inpatients.

5. Conclusion

Estimating 24-hour CCr by using algebraic equations may be misleading in aged patients. We demonstrated that ANN approach exceeded in performance than algebraic equations and ANN can be a feasible alternative for 24-hour CCr estimation in elderly inpatients. With the assistance of friendly software, it is not difficult to take advantages of artificial intelligence. We believe that our study using ANN in predicting 24-hour CCr will be an infrastructure for further investigations.

References

[1] Lindeman RD, Tobin J, Shock NW. Longitudinal studies on the rate of decline in renal function with age. J Am Geriatr Soc 1985; 33: 278-85.

[2] Du BD, Du Bois EF. A formula to estimate the approximate surface area if height and weight be known. Nutrition 1989; 5: 303-11.

[3] Cockcroft DW, Gault MH. Prediction of creatinine clearance from serum creatinine. Nephron 1976; 16: 31-41.

[4] Jelliffe RW. Creatinine clearance: bedside estimate. Ann Intern Med 1973; 79: 604-05.

[5] Levey AS, Bosch JP, Lewis JB, Greene T, Rogers N, Roth D. A more accurate method to estimate glomerular filtration rate from serum creatinine: a new prediction equation. Modification of Diet in Renal Disease Study Group. Ann Intern Med 1999; 130: 461-70.

[6] Hunter A, Kennedy L, Henry J, Ferguson I. Application of neural networks and sensitivity analysis to improved prediction of trauma survival. Comput Methods Programs Biomed 2000; 62:11-9.

[7] Krouwer JS, Monti KL. A simple, graphical method to evaluate laboratory assays. Eur J Clin Chem Clin Biochem 1995; 33: 525-27.

[8] Herman RA, Noormohamed S, Hirankarn S, Shelton MJ, Huang E, Morse GD, Hewitt RG, Stapleton JT. Comparison of a neural network approach with five traditional methods for predicting creatinine clearance in patients with human immunodeficiency virus infection. Pharmacotherapy. 1999; 19: 734-40. [9] Gabutti L, Ferrari N, Mombelli G, Marone C. Does cystatin C improve the precision of Cockcroft and