Non-invasive diagnosis of risk in dengue patients using bioelectrical impedance analysis and artificial neural network

Transcription

1 Med Biol Eng Comput (2010) 48: DOI 10.7/s z ORIGINAL ARTICLE Non-invasive diagnosis of risk in dengue patients using bioelectrical impedance analysis and artificial neural network F. Ibrahim T. Faisal M. I. Mohamad Salim M. N. Taib Received: 28 May 2009 / Accepted: 9 July 2010 / Published online: 4 August 2010 Ó International Federation for Medical and Biological Engineering 2010 Abstract This paper presents a new approach to diagnose and classify early risk in dengue patients using bioelectrical impedance analysis (BIA) and artificial neural network (ANN). A total of 223 healthy subjects and 207 hospitalized dengue patients were prospectively studied. The dengue risk severity criteria was determined and grouped based on three blood investigations, namely, platelet (PLT) count (less than or equal to 30,000 cells per mm 3 ), hematocrit (HCT) (increase by more than or equal to 20%), and either aspartate aminotransferase (AST) level (raised by fivefold the normal upper limit) or alanine aminotransferase (ALT) level (raised by fivefold the normal upper limit). The dengue patients were classified according to their risk groups and the corresponding BIA parameters were subsequently obtained and quantified. Four parameters were used for training and testing the ANN which are day of fever, reactance, gender, and risk group s quantification. Day of fever was defined as the day of fever subsided, i.e., when the body temperature fell below 37.5 C. The blood investigation and the BIA data were taken for 5 days. The ANN was trained via the steepest descent back propagation with momentum algorithm using the log-sigmoid transfer function while the sum-squared error was used as the network s performance indicator. The best ANN architecture of (3 inputs, 6 neurons in the hidden layer, and 1 output), learning rate of 0.1, momentum constant of 0.2, and F. Ibrahim (&) T. Faisal M. I. Mohamad Salim Department of Biomedical Engineering, Faculty of Engineering, University of Malaya, Kuala Lumpur, Malaysia fatimah@um.edu.my M. N. Taib Faculty of Electrical Engineering, Universiti Teknologi Mara, Shah Alam, Selangor, Malaysia iteration rate of 20,000 was pruned using a weight-eliminating method. Eliminating a weight of 0.05 enhances the dengue s prediction risk classification accuracy of 95.88% for high risk and.83% for low risk groups. As a result, the system is able to classify and diagnose the risk in the dengue patients with an overall prediction accuracy of.27%. Keywords Dengue fever Risk Severity Classify Bioelectrical impedance analysis Artificial neural network 1 Introduction Dengue disease is one of the current major public health affairs and is endemic in the Americas, southern Europe, North Africa, the eastern Mediterranean, Asia, Australia, various islands in the Indian Ocean, the south and the central pacific and the Caribbean [12, 29, 41]. Approximately, 2.5 billion people facing risk and it are predicted that this number will increase as transmission spreads to neighboring geographic regions [11, 13]. The World Health Organization (WHO) estimated of million cases of dengue fever (DF) occurs annually [11]. In Malaysia, DF was reported as early as 11 and the first report of the DF with hemorrhagic symptoms was made in 12 [30, 34, 36]. In 1973, the first major outbreak of dengue hemorrhagic fever (DHF) (45 cases) was reported [39]. Since that time, the increase of dengue cases in Malaysia has been very significant, e.g., 19,544 cases with 50 deaths reported in 1997, and then increased to 27,373 dengue cases with 58 deaths in 19 [1, 9, 34]. Recently, from 2004 to 2008, the number of dengue cases reported in Malaysia has increased from 33,895 (102 deaths) to 49,335 (112 deaths) [28]. Dengue virus is considered as arthropod-borne virus due its transmission cycle between Aedes mosquitoes and

2 1142 Med Biol Eng Comput (2010) 48: humans [41]. Dengue virus consists of four serotypes (DEN1, DEN2, DEN3, and DEN4) which are causing the DF, DHF, and dengue shock syndrome (DSS). Typically, DF begins with a sudden temperature increase accompanied by headache, myalgia, macular rash, loss of appetite nausea, vomiting, abdominal pain, metallic taste of food, change in psychological state, and moderate thrombocytopenia [2, 20, 22, 41]. If early clinical management or appropriate fluid therapy was not provided, DF will progress to DHF. The progress of DHF begins when the fever subsided or known as defervescence of the fever. DHF is an infection associated with an increase in microvascular permeability, a decrease in plasma volume, and in severe forms hypotension and shock. If the appropriate therapy still not provided, circulatory failure will occur and lead to the DSS. DSS is fatal stage manifested by rapid and weak pulse and narrow pulse pressure [41]. Therefore, delay in the fluid therapy management will lead to progress of the dengue disease to the maximum stage (DSS) and may cause fatality to the dengue patient. Accurate diagnosis in time and monitoring of severity of any dengue infection is needed in order to identify the severity of the disease and providing the appropriate treatment. In order to treat and control the dengue disease, many strategies have been developed and promoted. Two conventional techniques have been used to diagnose and to monitor the risk in DHF patients. The first technique is observing the onset and progression of plasma leakage by measuring the total increase in hematocrit (HCT) or the hemoglobin (Hb) concentration [15, 19, 20, 41]. The advantages of this method not only diagnose the DF but it can distinguish the DHF [14, 32, 41]. The second technique is to monitor the dengue patients platelet (PLT) counts and liver function [28]. Even though those techniques can give accurate diagnoses, they are time consuming, invasive, and may harm the patients [20, 21]. The reasons behind this as follow: those techniques require frequent blood taking which may cause further injury to the subcutaneous tissue and potentially hazardous to the DHF patients. Moreover, monitoring of the patients can be done by admitting and hospitalizing the patients. However, admitting and hospitalizing the patient cannot be arbitrarily or based on uncertain identification due to the huge number of the dengue patients in the country [20]. Recently, several studies were conducted to achieve accurate diagnosis without facing the above-mentioned drawbacks. Ibrahim et al. [20] applied bioelectrical impedance analysis (BIA) technique to monitor and classify the daily risk in DHF patients. Significant outcome was attained by using this technique and proved that the capability of the BIA to classify the daily risk of DHF patients. Another study [21] has reveals the non-invasive system for predicting the day of defervescence of fever in dengue patients using artificial neural network (ANN). Clinical symptoms and signs are used as inputs to the ANN. Accordingly, % prediction accuracy was achieved for predicting the day of defervescence of fever in dengue patients. Recently, Faisal et al. [6] employed self organized map clustering technique instead of the conventional statistical method to identify the risk criteria for classifying risk in the dengue patients. This paper focuses on the diagnosis of risk classification in the dengue patients utilizing the BIA and a multilayer feed-forward neural network (MFNN) techniques. 2 Methodology The procedures for designing the dengue risk diagnose system shows in Fig Clinical data Database comprises of 223 healthy subjects (158 females and 65 males) and 207 ( females and 115 males) serologically [5, 27] confirmed dengue patients during their hospitalization were prospectively studied. The dengue database was divided into blood investigation and BIA data. The blood investigation consists of 27 parameters such as PLT, HCT, aspartate aminotransferase (AST), alanine aminotransferase (ALT), and Hb [18, 20]. The blood investigation and the BIA data were taken for 5 days with reference to the day the fever subsided. Fever day?0 is defined as the day the fever subsided, i.e., when the body temperature fell below 37.5 C. Fever days after fever day?0 are fever days?1 and onwards [18, 20]. The blood measurement was taken from fever day?0 until the fever day?4 (fourth day after fever subsided). The BIA database comprises of patient information and 17 BIA parameters BIA data reactance X c Gender Clinical Data Determine the day of fever Normalize the data (0-1) Training MFNN and optimization Pruning the ANN Testing Final ANN Risk quantification Risk groups Fig. 1 The procedures of designing the dengue risk diagnose system

3 Med Biol Eng Comput (2010) 48: including the resistance, reactance, body capacitance (BC), fat mass, intracellular water, etc. 2.2 BIA BIA is an in vivo technique involved the application of a small average constant current of less than 1 ma at a single frequency of 50 khz through the human body, and measuring the body s bioelectrical tissue conductivity (BETC) parameters, namely, R, phase angle (a), BC, and capacitive reactance (X c ) via four surface electrodes [35]. Previous studies [19, 20, 22] shown that single frequency technique is able to give good results in dengue patients. Two electrodes were placed on the patient s right hand, one at the base of the knuckles and another slightly above the wrist joint. Another two electrodes were placed on the right foot; one nears the base of the toes and the other slightly above the ankle joint. A constant current was applied to the base of the knuckles and base of the toes, and the voltage signal was picked up by the other two sensor electrodes (slightly above the ankle and wrist joint). The voltage drop will determine the resistance and reactance (X c ) of the whole body. This data can then be converted to estimate the extracellular water (ECW), intracellular water, fat free mass (FFM), and fat mass through regression equations. 2.3 Risk quantification In this study, the severities of dengue risk criteria were determined based on the following blood investigations [18, 20, 41]: i. PLT count is less than or equal to 30,000 cells per mm 3 [4, 18] ii. HCT increase by more than or equal to 20% [41] iii. AST and ALT levels rose by fivefold the normal upper limit for AST and ALT [18, 26]. The risk quantification was performed on daily basis where blood investigations of the patients were evaluated for each day. Based on the blood investigation, the patients were then divided into two groups [18, 20]: i. Group 2 (Lower risk group) accounted for the DHF patient who did not experienced any of the defined risk criteria, or experienced only 1 of the 3 risk criteria. ii. Group 3 (Higher risk group) accounted for DHF patient who experienced 2 or more risk criteria. The dengue patients were then classified according to their groups and subsequently, their corresponding BIA data were obtained and quantified. The healthy subjects were automatically grouped as the control group (Group 1) with no past medical history and no blood investigations evaluated. 2.4 Pre-processing of ANN Database preparation involves parameter selection, data massaging, and data grouping Parameter selection In the parameter selection phase from the BIA database only three parameters were selected to be the input of the MFNN. The input parameters are: day of fever, gender, and reactance values. The selection of these parameters was based on multi-logistic analysis [18]. The quantified risk was assigned as the network target Data massaging Before the data is fed into the MFNN, it was first massaged according to the data type. Data massaging involves restructuring the range of neural network input and output values between the ranges of 0 1. Massaging is done due to the fact that neural network works best when all its input and output vary within the range of 0 1 [31]. Generally, massaging data can be performed by using a general equation which is: Massage value Actualinputvalue Minimuminputvalue ¼ Maximum input value Minimum input value Therefore, massaging the data for day of fever will be 0, 0.25, 0.5, 0.75, and 1 for days 0, 1, 2, 3, and 4, respectively. The massaging data for risk group are 0, 0.5, and 1, for control data, low risk, and high risk, respectively. Eventually, 0 and 1 are the massage data for the female and male, respectively Data sampling Data comprises of 223 healthy subjects and 207 dengue patients were arranged randomly into the training and testing in the ratio of 70:30 as distributed in Table 1. The total number of training and testing data was 303 and 127, respectively. Since, the fever days are from days 0 4, each patient will have five samples and the total samples are Table 1 Data distribution for training and testing the ANN Groups Training Testing Female Male Female Male Control data Dengue patients Total Total samples

4 1144 Med Biol Eng Comput (2010) 48: ( ) and 635 ( ) for training and testing data, respectively. However, many of the patients have early recoveries and discharged before the fever days 4. Therefore, a unique data sampling of 1195 training and 509 testing were obtained. 2.5 ANN In biomedical field, ANN has been utilized in varies application such as modeling, data analysis, and diagnostic classification [3, 16, 38]. The most common ANN model used in clinical medicine is the Multilayer Perceptron (MLP) [25]. The most widely used connection pattern in ANN is the three layer back propagation neural network which has been proved to be useful in modeling input output relationship [31] while the most commonly used transfer functions are linear, log-sigmoid and tan-sigmoid functions [37] Development of MFNN This study involves in the development of a MFNN with three inputs. It is trained using the steepest descent with momentum back propagation algorithm in Matlab environment. The measurement of MFNN performance was observed by using the sum-squared error (SSE) and total prediction accuracy of network to the testing data. Training is best when the MFNN is capable to achieve lowest SSE value [3, 16 18, 25]. The network was trained using the Traingdm algorithm with Logsig and Purelin transfer functions. The optimum network architecture was systematically determined by varying the training parameters (the number of neuron in the hidden layer, learning rate, momentum constant, and iteration rate). As one parameter was being varied to find its optimum value, the other parameters were kept constant. The optimum value of each parameter is determined by the two performance indicators that were assigned that are SSE and total prediction accuracy of the network Pruning the ANN Pruning the ANN not only increases the speed of the system and improve generalization [10] but it can also increase the accuracy of the diagnosis [21]. Three methods have been used to prune the neural network, weight-eliminating cost function [40], sensitivity calculation [17], and magnitude based method [25]. In this study magnitude based method was used. In the magnitude based method, the weight with the smallest magnitude will be removed from the network. The decision of which, weight will be removed, is based on observing the accuracy of the network when the specific weight is removed. Accordingly, the procedure of pruning the ANN as follows: I. Select the weight with minimum magnitude from the best network and set it as a threshold. II. Set the threshold for weights from hidden units to output unit and from the input to the hidden unit. III. Remove all weights equal to the threshold. IV. Check the accuracy of the system. V. If the accuracy is lower stop and save the previous ANN weights as a final weight. VI. If the accuracy of the ANN is same or higher move to step VII. VII. Choose new threshold which is great magnitude than the previous one. VIII. Set the threshold for weights from hidden units to output unit and from the input to the hidden unit IX. Remove all the weight below the threshold. X. Repeat step IV. 3 Results 3.1 Risk quantification The dengue patients blood investigation results were classified into as high risk and low risk groups as shown in Table Results for determination the optimum MFNN Number of neuron in the hidden layer The number of neuron in the hidden layer was varied from 1 to 19 while other parameters including the learning rate, momentum constant, and training epochs were fixed to 0.1, 0.7, and 20000, respectively. Figure 2 shows the values of the SSE and the total accuracy when the hidden layer size is varied. It can be seen that networks with hidden layer sizes 11, 17, and 19 produce SSE values lower than The total accuracies for these networks are only.73, 91.55, and 88.21%, Table 2 Result of blood investigations for risk classification in dengue patients Fever days Number of patients Group 2 (low risk) Group 3 (high risk) Female Male Female Male

5 Med Biol Eng Comput (2010) 48: respectively, which is relatively low. However, the accuracy of network with hidden layer sizes 6 produces the highest total accuracy of 95.48% and acceptable SSE value of Hence, hidden layer size of 6 was chosen the optimum hidden layer size. Although the network s SSE that was chosen did not meet the error goal which is set to 0.01, the error is low enough since the network input is set to only one decimal point Learning rate The MFNN learning rate was varied from 0.1 to 0.9. Number of neuron in the hidden layer was set at the optimal value of 6, which was obtained during the optimal hidden neuron determination. Other parameters such as momentum constant and training epochs were set to 0.7 and 20,000, respectively. Figure 3 shows the overall network performance for various learning rates. The Figure depicts that the learning rate of 0.6 produces the lowest SSE value of and low Number of Nerons in the hidden layer vs Sum Squared Error (SSE) and the Total Accuracy (%) Total Accuracy (%) Sum Squared Error (SSE) Fig. 2 Plot of SSE and the total accuracy against the number of hidden neurons total prediction accuracy of 91.16%. Since the total accuracy of a network is more crucial in this application, compensation is done by selecting network with learning rate of 0.1 with the highest total prediction accuracy and adequate SSE value of The prediction accuracy of the network is 95.48%, an increase of 4.32% in prediction accuracy with difference of in SSE value. Hence, learning rate of 0.1 was chosen as the optimum learning rate Momentum constant The momentum constant was varied from 0.1 to 0.9. Training epochs was set to 20,000 while the number of neuron in the hidden layer and the learning rate were set to the optimal values of 6 and 0.1, respectively. Figure 4 indicates the momentum constant 0.7 produces the lowest SSE value of However, the prediction accuracy is only 95.48%. On the other hand, momentum constant of 0.2 has the highest prediction accuracy. The prediction accuracy and SSE of network with momentum constant 0.2 are.86% and 0.019, respectively. The network produces an increment of 1.38% of total accuracy with only differences in SSE value. Hence, momentum constant of 0.2 was chosen as the optimum value for the MFNN Training epochs The final step for the determination of optimum MFNN is to find the best training epochs or iteration rates. The training epochs was varied from 10,000 to 50,000 with a constant increment of 10,000. Other parameters such as hidden layer size, learning rate, and momentum constant were set at the optimal values. Figure 5 illustrates that training epochs of 10,000 produces the highest SSE values and the lowest accuracy. This Learning rate vs Sum Squared Error (SSE) and the Total Accuracy (%) Total Accuracy (%) Sum Squared Error (SSE) Fig. 3 Plot of SSE and the total accuracy against the learning rate Momentum constant vs Sum Squared Error (SSE) and the Total Accuracy (%) Total Accuracy (%) Sum Squared Error (SSE) Fig. 4 Plot of SSE and the total accuracy against the momentum constant

6 1146 Med Biol Eng Comput (2010) 48: Iteration rate vs Sum Squared Error (SSE) and the Total Accuracy (%) Total Accuracy (%) 0.03 Sum Squared Error (SSE) indicates that the training epochs is too short to allow the network to converge and leads to the low prediction accuracy. However, as the epochs values increase, the SSE values reduces, leading to higher prediction accuracy. It can be shown that networks with training epochs of 50,000 produces the lowest SSE values of However, total prediction accuracy for the network was 93.12% which is relatively low as compared to the total prediction accuracy given by network with training epochs of 20,000. Thus, the training epochs of 20,000 with the highest prediction accuracy of.86% and acceptable SSE of was chosen for the MFNN. This network gives an increment of 1.38% in total prediction accuracy with only differences in SSE value. Hence, the optimum MFNN that is used for the risk prediction in dengue patients is by network architecture of (3 network inputs, 6 neurons in the hidden layer, one network output), learning rate of 0.1, momentum constant of 0.2, and training epochs of Pruning the MFNN Fig. 5 Plot of SSE and the total accuracy against the training epochs The final optimum MFNN obtained was pruned to enhance the performance of the network. The smallest weight of 0.02 was eliminated from the optimized MFNN. The accuracy comparison results before and after pruning the MFNN is shown in Fig. 6. Figure 6 illustrates that eliminating the 0.02 weight has no effect in the MFNN network, however, it has increased the processing rate of the network. The results for eliminating any weight less than or equal to the 0.05 weight is shown in Fig. 7. The pruning process has reduced the control group prediction accuracy to 95.81%. On the other hand, the prediction accuracies for the low and high risk groups have improved to.83 and 95.88%, respectively. Figure 8 depicts the pruning results for the weights less than or equal 0.1. The prediction accuracies of the MFNN Accuracy % ANN AFTER PRUNING Threshold pruning ANN BEFORE PRUNING Control Data Low risk High risk Over all Fig. 6 Comparison results before and after pruning the MFNN at the weight of 0.02 Accuracy % Dengue Diagnostic System (0.05) ANN AFTER PRUNING ANN BEFORE PRUNING Control Data Low risk High risk Over all Fig. 7 Comparison results before and after pruning the MFNN at the weight of 0.05 for the control and low risk groups before pruning are and 95.93%, respectively. However, pruning the MFNN has reduced the prediction accuracies to.15 and.5% for control and low risk groups, respectively. The prediction accuracy for high risk group before and after pruning the MFNN are.78 and.85%, respectively. Due to the reduction of the accuracy in the control and low risk groups, it has affected the overall prediction accuracy of the MFNN system to 91.% which is significantly lower than the network before being pruned.86%. Thus, any further changes by eliminating weights higher than 0.05 will affect the performance of the system. Accuracy % Dengue Diagnostic System (0.1) ANN AFTER PRUNING ANN BEFORE PRUNING Control Data Low risk High risk Over all Fig. 8 Comparison results before and after pruning the MFNN at the weight of 0.1

7 Med Biol Eng Comput (2010) 48: Consequently, it can be concluded that pruning the network by eliminating all the weights less than or equal to 0.05 improves the prediction accuracy of the MFNN for the high risk and low risk groups in dengue patients. 4 Discussions In this study,.86% overall prediction accuracy for diagnosis of risk in dengue patients using BIA and ANN has been achieved. The diagnostic accuracy for diagnosis the control, high risk, and low risk groups were,.78, and 95.93%, respectively. However, when the pruning technique were implemented, the overall prediction accuracies were.27% with 95.88,.83, and 95.81% for high risk, low risk, and control groups, respectively. Therefore, pruning the ANN reduced the overall prediction and control group accuracies by 0.59 and 4.19%, respectively, however, it is favorable to prune the ANN since the diagnosis accuracy of the low risk and high risk groups have been increased. Thus, this work and recent findings [7, 8] have shown that the noninvasive intelligent system and neural network are promising techniques to predict risk in dengue patients. However, these results of single bioimpedance measurements is limited to the extracellular cell and can be enhanced and improved by using bioimpedance spectroscopy techniques [23, 24, 33] where measurement can be obtained in the intracellular cell at different frequencies ranges. Acknowledgments This research was supported by University of Malaya, Sultan Iskandar Johore Foundation, and Ministry of Science, Technology and Innovation (MOSTI) (E-Science Fund ). References 1. Annual Report, WHO Collaborating and Research (19) Dengue haemorrhagic fever. Department of Medical Microbiology, Kuala Lumpur, Malaysia 2. Balmaseda A et al (2006) Serotype-specific differences in clinical manifestations of dengue. Am J Trop Med Hyg 3(74): Basheer IA, Hajmeer M (2000) Artificial neural networks: fundamentals, computing, design, and application. J Microbiol Methods 43: Chua MN, Molanida R, De Guzman M, Laberiza F (1993) Prothrombin time and Partial thromboplastin time as a predictor of bleeding in patient with DHF. Southeast Asian J Trop Med Public Health 5. Chungue E, Boutin JP, Roux J (19) Significance of IgM titration by an immunoenzyme technic for the serodiagnosis and epidemiological surveillance of dengue in French Polynesia. Res Virol 3(140): Faisal T, Taib MN, Ibrahim F (2010) Reexamination of risk criteria in dengue patients using the self-organizing map. Med Biol Eng Comput 48: Faisal T, Taib MN, Ibrahim F (2010) A noninvasive intelligent approach for predicting the risk in dengue patients. Expert Syst Appl 37: Faisal T, Taib MN, Ibrahim F (2010) Neural network diagnostic system for dengue patients risk classification. J Med Syst (in press) 9. Fang R et al (12) The dengue epidemic in Malaysia. Epidemiological, serological and virological aspects. Southeast Asian J Trop Med Public Health 14: Goh YS, Tan EC (19) Pruning neural networks during training by backpropagation. TENCON. IEEE region 10 s ninth annual international conference 11. Gubler DJ (19) Dengue and dengue hemorrhagic fever. Clin Microbiol Rev 11: Gubler DJ (2002) Epidemic dengue/dengue hemorrhagic fever as a public health, social and economic problem in the 21st century. Trends Microbiol 10: Hales S, de Wet N, Maindonald J, Woodward A (2002) Potential effect of population and climate changes on global distribution of dengue fever: an empirical model. Lancet 360: Halstead BS (19) Antibody, macrophages, dengue virus infection, shock, and hemorrhagic: a pathogenetic cascade. Rev Infect Dis 11: Harris E et al (2000) Clinical, epidemiologic, and virologic features of dengue in the 19 epidemic in Nicaragua. Am J Trop Med Hyg 63: Haselsteiner E, Pfurtscheller G (2000) Using time-dependent neural networks for EEG classification. IEEE Trans Rehabil Eng 8: Haykin S, Neural networks (19) A comprehensive foundation. Macmillan, New York 18. Ibrahim F (2005) Prognosis of dengue fever and dengue hemorrhagic fever using bioelectrical impedance. Ph.D. Thesis, Department of Biomedical Engineering, Faculty of Engineering, University of Malaya 19. Ibrahim F, Ismail NS, Taib MN, Wan Abas WAB (2004) Modeling of hemoglobin in dengue fever and dengue hemorrhagic fever using bioelectrical impedance. Physiol Meas 25: Ibrahim F, Taib MN, Wan Abas WAB, Chan CG, Sulaiman S (2005) A novel approach to classify risk in dengue hemorrhagic fever (DHF) using bioelectrical impedance. IEEE Trans Instrum Meas 54(1): Ibrahim F, Taib MN, Wan Abas WAB, Chan CG, Sulaiman S (2005) A novel dengue fever (DF) dengue and hemorrhagic fever (DHF) analysis using artificial neural network. Computer Methods Programs Biomed 79: Ibrahim F, Taib MN, Wan Abas WAB, Chan CG, Sulaiman S (2008) A new approach to classify risk in dengue infection using bioelectrical impedance analysis (BIA). World Health Org Dengue Bull 31: Jaffrin MY, Morel H (2009) Extracellular volume measurements using bioimpedance spectroscopy-hanai method and wrist-ankle resistance at 50 khz. Med Biol Eng Comput 47: Jaffrin MY, Fenech M, Moreno MV, Kieffer R (2006) Total body water measurement by a modification of the bioimpedance spectroscopy method. Med Biol Eng Comput 44(10): John T, Wei ZZ, Barnhill SD, Madyastha R (19) Understanding artificial neural networks and exploring their potential applications for the practicing urologist. Urology 52: Kuo CH, Tai DI, Chang-Chien CS, Lan CK, Chiou SS, Liaw YF (19) Liver biochemical test and dengue fever. Am J Trop Med Hyg 3(47): Lam SK, Devi S, Pang T (17) Detection of specific IgM in dengue infection. Southeast Asian J Trop Med Public Health 4(18): Ministry of Health Malaysia (2009). MohPortal/newsFull.jsp?action=load&id=401. Assessed 5 Jul 2009

8 1148 Med Biol Eng Comput (2010) 48: Monath TP (19) Dengue: the risk to developed and developing countries. Proc Natl Acad Sci USA 91: More FW (14) Observations on dengue fever in Singapore. J Malaya Branch Br Med Assoc 1: Negnevitsky M (2002) Artificial intelligence, a guide to intelligent system. First Edition, Pearson education. ISBN: Nimmannitya S (17) Clinical spectrum and management of dengue haemorrhagic fever. Southeast Asian J Trop Med Public Health 18: Paterno AS, Stiz RA, Bertemes-Filho P (2009) Frequencydomain reconstruction of signals in electrical bioimpedance spectroscopy. Med Biol Eng Comput 47(10): Rebecca George MD (19) Current status of the knowledge of dengue/dhf/dss in Malaysia: clinical aspects. 15th Annual Convention of the Philippine Society foe Microbiology and Infectious Diseases 35. Rigaud B, Morucci JP, Chauveau N (19) Bioelectrical impedance techniques in medicine part I: bioimpedance measurement second section: impedance spectrometry. In: Bourne JR (ed) Critical reviews in biomedical engineering. Begell House, New York, vol Rudnick A et al (15) Mosquito borne haemorrhagie fever in Malaysia. Br Med J 1: Sinha SK, Fieguth PW (2005) Projection neural network model for classification of pipe defects. J Autom Constr 15(1): Sun M, Sclabassi RJ (2000) The forward EEG solutions can be computed using artificial neural networks. IEEE Trans Biomed Eng 47: Wallace Hazel G et al (10) Dengue haemorthagic fever in Malaysia 1973 epidemic. Southeast Asian J Trop Med Public Health 11: Weigend AS, Rumelhart DE, Huberman BA (1991) Generalization by weight elimination with applications to forecasting. In: Lippmann R, Moody J, Touretzky D (eds) Advances in neural information processing 3, pp World Health Organization (1997) Dengue haemorrhagic fever: diagnosis, treatment, prevention and control, 2nd edn. WHO, Geneva