Abstract
This paper aims to construct intelligence models by applying the technologies of artificial neural networks including back-propagation network (BPN), generalized feedforward neural networks (GRNN), and modular neural network (MNN) that are developed, respectively, for the early detection of chronic kidney disease (CKD). The comparison of accuracy, sensitivity, and specificity among three models is subsequently performed. The model of best performance is chosen. By leveraging the aid of this system, CKD physicians can have an alternative way to detect chronic kidney diseases in early stage of a patient. Meanwhile, it may also be used by the public for self-detecting the risk of contracting CKD.
1. Introduction
According to the statistical data announced by the Department of Health of Taiwan’s government in 2010 [1], the mortality caused by kidney disease has been ranked in the 10th place in all causes of death in Taiwan and thousands of others are at increased risk. The mortality caused from kidney disease is estimated as 12.5 in every 100,000 people. As a result, it costs as high as 35 percent of health insurance budget to treat the chronic kidney disease (CKD) patients with the age over 65 years old and end-stage kidney disease patients in all ages. It occupies a huge amount of expenditures in national insurance budget.
Regarding the measurement of serious levels of CKD, presently glomerular filtration rate (GFR) is the most commonly measuring indicator used in health institutions to estimate kidney health function. The physician in the health institution can calculate GFR from patient’s blood creatinine, age, race, gender, and other factors depending upon the type of formal-recognized computation formulas [2, 3] employed. The GFR may indicate the health of a patent’s kidney and can also be taken to determine the stage of severity of a patient with or without kidney disease.
In this paper, we aim to develop a feasible intelligent model for detecting CKD for evaluating the severity of a patient with or without CKD. The input data for model development and testing is collected from the health examination which is periodically carried out by the collaborative teaching hospital of this research.
2. The Major Methods for Measuring Chronic Kidney Disease
As it is mentioned in prior section, the GFR is the most common method used to measure kidney health function. It refers to the water filterability of glomerular of people’s kidney. The normal value should be between 90 and 120 mL/min/1.73 m2 (i.e., measured by mL per minute per 1.73 m2). There are three common computation methods of GFR, which are (1) removing rate of 24-hour urine creatinine (i.e., Creatinine Clearance Rate, Ccr), (2) Cockcroft-Gault formula (also known as C-G formula), and (3) Modification of Diet in Renal Disease formula (also known as MDRD) [4]. The CKD is categorized into five stages by making use of GFR to measure kidney function. Although the course of the change from stage one to five may usually last for years, it sometimes may enter into fifth stage pretty soon resulted in the necessity of dialysis or kidney transplant.
Again, National Kidney Foundation’s Kidney Disease Outcomes Quality Initiative (KDOQI) [3] provides a conceptual framework for the diagnosis of the severity stages of CKD based on the different function levels of glomerular filtration rate (GFR). The new system represented a significant conceptual change, since kidney disease historically had been categorized mainly by causes. The diagnosis of CKD relies on markers of kidney damage and/or a reduction in GFR. Stages 1 and 2 define conditions of kidney damage in the presence of a GFR of at least 90 mL/min/1.73 m2 or 60 to 89 mL/min/1.73 m2, respectively, and stages 3 to 5 define conditions of moderately and severely reduced GFR irrespective of markers of kidney damage. The summary of this guideline is shown in Table 1 [3]. However, Levey et al. [5] especially mentioned that although this guideline was endorsed by the Kidney Disease: Improving Global Outcomes (KDIGO) in 2004 and this framework was constantly promoted to increase the attention to chronic kidney disease in clinical practice, research, and public health, it had also generated debate. It is the position of KDIGO and KDOQI that the definition and classification should reflect patient prognosis and that an analysis of outcomes would answer key questions underlying the debate. The common definition of CKD has facilitated comparisons between studies. Nevertheless, there are limitations to this classification system, which is by its nature simple and necessarily arbitrary in terms of specifying the thresholds for definition and different stages. When the classification system was developed in 2002 [3], the evidence base used for the development of this guideline was much smaller than the CKD evidence base today. Therefore, this guideline has been constantly revised from then on.
In Taiwan, the Taiwan Society of Nephrology (TSN) also presented the self-detecting method of kidney for the public. At present, the MDRD formula [3, 6] is recognized as a mostly common method adopted by kidney physicians to estimate GFR from serum creatinine level. Therefore, in this paper we take MDRD to calculate the GFR for the detection. The method of computational formula is shown in formula (1). The input data for the GFR calculation of each individual case in health examination are provided by the collaborative teaching hospital. We also used the calculated results as the desired (targeted) value to develop our neural network models. One has Note: For female the result should be multiplied by a factor of 0.742.
It is specially worthy of note that the recent study from Matsushita et al. [7] indicates that although the Modification of Diet in Renal Disease (MDRD) study equation is recommended for estimating GFR, the Chronic Kidney Disease Epidemiology Collaboration (CKD-EPI) has proposed an alternative equation, which is known as CKD-EPI. The CKD-EPI applies different coefficients to the same 4 variables which include age, sex, race, and serum creatinine level, used in the MDRD study equation. The study takes the data from more than one million participant cases residing in 40 countries or regions. They find that the CKD-EPI equation estimates measured GFR more accurately than the MDRD study equation in most of the study areas. It shows approximately one-fourth of cases were reclassified to a higher estimated GFR category by the CKD-EPI equation compared with the MDRD study equation. In this one-fourth of cases are reclassified upward in GFR figure by CKD-EPI, 24.4% in the general population cohorts, 15.4% in the high risk cohorts, and 6.6% in the CKD cohorts. This improvement by CKD-EPI classification may lower the prevalence of CKD. Participant cases who are reclassified upward had lower risks of mortality and end-stage renal disease (ESRD) compared with those not reclassified [7].
3. Artificial Neural Network
The artificial neural network (ANN), usually simply called neural network (NN), is a mathematical model or computational model that is inspired by the structure and/or functional aspects of biological neural networks. Kriesel [6] indicated neural networks were a bioinspired mechanism of data processing that enables computers to learn technically similar to human-being brain. A neural network consists of an interconnected group of artificial neurons, and it processes information using a connectionist approach to computation. In most cases, an ANN is an adaptive system that changes its structure based on external (input) or internal information that flows through the network during the learning phase. It is properly the most prestigious and adoptable model in all application models in the field of artificial intelligence.
There are many different types of neural network models derived from the generic structure of ANN. In this paper’s study, three neural network models employed for the experiment and comparison include the back-propagation neural network (BPNN) [6, 8], the generalized feedforward neural network (GFNN) [9, 10], and the modular neural network (MNN) [11]. The generic architectures of three-layer back-propagation neural network, generalized feedforward neural network and modular neural network are illustrated in Figures 1, 2, and 3, respectively.
4. Research Materials and Methods
Based on the literature reviews and expert interviews, this paper takes three neural network models including back-propagation neural network, feedforward neural network, and modular neural network, respectively, to generate the detection model for chronic kidney disease. The key influence input factors for each model are determined and verified from professional experts and physicians of kidney diseases. The set of influence factors derived from GFR computation formulas is shown in Table 2 and another set of key influence factors selected and determined in this paper as the input of neural networks is shown in Table 3, respectively. The classification performances of two input sets for model development are compared.
The input data set for developing neural networks are collected from the cases of health examination provided by the collaborative hospital of this study. 1161 heath examination cases covering past few years are selected. Before they are input for training and testing network models, a data preprocessing is processed to remove duplication and correct the error, inconsistency, and missing fields in each case record. Among these cases, there exit some unknown duplication cases which have unknown causes, they are identified and removed. Some cases have error in certain fields such as the figure being beyond its reasonable range or in a questionable high or low level. This is often seen in the figures for some physiological test. We try to confirm these figures with the authority of health examination center of hospital and try to correct it, otherwise they are removed. Some cases show postal code error or appear inconsistent with the personal contact address or vice versa, using different representation to refer the same meaning, which is commonly seen in gender filed and name field. More significantly, because not every subject conducts a complete health examination, certain fields used for risk measurement are missing. These cases are simply removed as well because they can not be used for model development. By this process, we can ensure the accuracy, completeness, and integrity of input data. After data preprocessing, only most accurate 430 patient cases remain for the development of intelligent models. Among these cases, 145 cases are prediagnosed as negative with CKD, and 285 cases are prediagnosed as positive. The details are shown in Table 4.
5. Model Development
The development for a specific network model can be viewed as a series of modeling and simulation. We attempt to pursue the feasible model through the experiment of different combinations of modeling parameters selected for different network models. The parameters selected for experiment include the learning rules, the transform functions, the number of hiding layers, the number of neurons in each hidden layer, the weight update methods, and the criteria for the determination in model training and testing. The well-known NBuilder of neural solution is the tool adopted for model development. Among 430 cases, 300 cases are selected for training, 100 cases are used for testing, and the rest of 30 cases are taken for cross-verification during the model development. This data distribution is shown in Table 4. The desired output of classification for each case is formally identified beforehand by the professional physicians of chronic kidney disease from the collaborative teaching hospital by using GFR computation formula. The hybrid model of combining each individual neural network and genetic algorithm (GA) is also conducted in this research. In other words, GA will be taken to combine with back-propagation neural network, feedforward neural network, and modular neural network, respectively, in model development and comparison as well.
5.1. The Model Development of Back-Propagation Neural Network
Through a series of training and testing with the different sets of parameter combination selected, we gain the best model for back-propagation neural network in terms of its respective network parameters with respect to two different sets of input factors. One set is by adopting the key factors used in computation formula (simply called Model PA hereafter). The other set is by adopting the key factors selected in this paper (simply called Model FB hereafter). Table 5 shows these two best settings of parameters. The classification accuracy, sensitivity, and specificity with respect to Model FA and FB, respectively, are shown in Table 6. In Table 6, the metric of “accuracy” is used to measure the classification of accuracy with the proportion of the sum of the number of true positives and the number of true negatives. The metric of “sensitivity” is used to measure the proportion of actual positives which are correctly classified as such. The metric of “specificity” is used to measure the proportion of negatives which are correctly identified. From the results shown in Table 6, a pretty good classification result is gained with BPN both for Model FA and FB, while it shows a significant drop in classification performance in model testing stage both in BPN and BPN plus GA. By the results shown in Table 6, pure BPN model may gain better results in accuracy measure, while BPN plus GA may gain better results in sensitivity measure. The result also shows the model with the adoption of the key factors used in computation formula gains better results in all three measures. These results show that a hybrid model with the combination of BPN and GA seams does not improve the model performance in accuracy measure but it is helpful to improve sensitivity measure.
5.2. The Model Development of Generalized Feedforward Neural Network
Table 7 shows the best network parameters settings for the development of generalized feedforward neural network (GFNN). The classification accuracy, sensitivity, and specificity with respect to Model FA and FB, respectively, are shown in Table 8. As the results in Table 8 show, pure GFNN obtains a perfect classification percentage in three measures including accuracy, sensitivity, and specificity in training stage while GFNN plus GA apparently may gain better results in testing. However, they also show the results gained from GFNN are not so good in all three measures as those gained from BPN and BPN plus GA, respectively.
5.3. The Model Development of Modular Neural Network
Table 9 shows the parameters for the test of modular neural network (MNN). The classification accuracy, sensitivity, and specificity with respect to Model FA and FB, respectively, are shown in Table 10. As the testing results in Table 10 show, pure MNN is the same as prior two models which may obtain a perfect classification percentage in three measures in training stage, but MNN plus GA apparently may gain worse results in testing stage, which is different from the results shown in prior two models.
6. Performance Comparison of Models
The performances of three neural network models developed for comparison in this paper in terms of detection (i.e., classification) accuracy measure are summarized and shown in Tables 11 and 12, respectively. Table 11 shows the detection accuracy in three pure neural network models while Table 12 shows the detection accuracy with GA model embedded in three respective pure models. We found BPN may gain the best accuracy regardless of in both Model FA and Model FB which are measured with 94.75% and 72.42% accuracy shown in Table 11, respectively in model testing. However, Model FA, which is the test by adopting the key factors used in computation formula, may gain much better performance than Model FB, which is the test by adopting the key factors selected in this paper. The same result is gained with the GA embedded in each fundamental model in test. As the results observed from Table 12, both BPN plus GA and GFNN plus GA gain close results in accuracy measure which is much better than MNN plus GA.
By further observation from the results of model test shown in Section 4, it is found that almost all models employed in test may gain near 100% accuracy in CKD detection in the training stage regardless of which sets of influence factors used in model training. However, if further model testing is conducted, it is found that the network models with the input of influence factors of CKD used by physicians employed in the computational formula always show better detection performance in all three aspects of measure including accuracy, sensitivity, and specificity measures. As we can see from Table 11, the BPN gains the highest 94.75% accuracy measure in the testing stage among three fundamental neural network models while GFNN gains only 86.63% in accuracy measure which is the lowest performance in three models.
Through further observations from test results as it is shown in Table 12, we find the hybrid network model of GFNN with GA embedded may significantly show improvement in detection performance in all three measures from its fundamental model in the testing stage although the reversed enhancement is shown in BPN and MNN. Again, by the test figures shown in Section 4, although pure neural network models with GA show degraded performance improvement for Model FA in testing stage in terms of three measures in accuracy, sensitivity, and specificity, it shows significant improvement for Model FB. As a result, we conclude a hybrid model indeed may improve detection performance. This result may conclude that GA provides no benefit in the yield of better detection performance with the adoption of the key factors used in computation formula as the input to all models in testing.
According to the detection performance shown in Tables 11 and 12, we conclude that BPN might be the best-fitting Model FA among three fundamental neural network models employed in detecting CKD while the models BPN and GFNN with GA embedded, respectively, might be the best hybrid models for CKD detection but only GFNN plus GA can gain enhancement in detection performance in both sets of influence factors as input. These conclusions should be further verified and compared with other models selected for test in later studies before they can be assured.
7. Conclusion
We conclude that neural network models developed for CKD detection may effectively and feasibly equip medical staff with the ability to make precise diagnosis and treatment to the patient.
In the future study, further model modification and testing for the intelligence models developed in this paper should be conducted to enhance the accuracy in detection performance and to ensure they are sufficiently good for being truly employed in medical practice. In the meantime, different intelligence models can be widely and persistently applied for system development in this domain application as well in order to search for a best one model to be adopted. In the future system development, it is worthwhile to deploy the system to the cloud platform so that the public users can also use this system to conduct a self-detection of having had CKD.