Dengue has become a major international public health concern. The incidence of dengue has grown dramatically around the world in recent decades. Over 2.5 billion people over 40% of the world’s population are now at risk from dengue. WHO currently estimates there may be 50–100 million dengue infections worldwide every year.
Dengue is transmitted to humans by the bite of Aedes mosquitoes. According to WHO (2014), “Dengue causes a severe flu-like illness, and sometimes a potentially lethal complication called dengue haemorrhagic fever”. About 2.5% of those that are infected by dengue die since dengue has neither treatment nor vaccination.Plasma leakage is the major cause of mortality and morbidity in patients with dengue hemorrhagic fever. So that early recognition of plasma leakage and prompt initiation of appropriate treatment are vital. There are only few researches which are done for plasma leakage detection in patients with DHF.
Dengue virus infections may be asymptomatic or may lead to undifferentiated fever, dengue fever (DF) or Dengue haemorrhagic fever (DHF) with plasma leakage that may lead to Dengue shock syndrome (DSS).
DF is generally an acute febrile illness, with severe headache, myalgia, arthralgia and rashes. Leucopenia and thrombocytopenia may also be observed. Although DF may be benign, it could be an incapacitating disease with severe headache, muscle and joint and bone pains. Occasionally unusual haemorrhage such as gastrointestinal bleeding, hypermenorrhea and massive epistaxis may occur.
Undifferentiated fever and classical dengue fever can be managed as any other viral fever with symptomatic treatment. However, often it is difficult to differentiate DF from DHF in the early phase (febrile phase) of the illness.
DHF is characterized by the acute onset of high fever and is associated with signs and symptoms similar to DF in the early febrile phase. Plasma leakage is the hallmark of DHF which occurs soon after the end of the febrile phase. There is a tendency to develop DSS due to plasma leakage. Therefore suspected DF and DHF patients should be closely monitored to identify patients with DHF.
The degree and the rate of plasma leakage in DHF can vary. It can be minimal in some patients while in others it can be very significant. The leak usually starts slowly, increases gradually, slows down and then ceases altogether at the end of leakage phase (usually within 48 hours from the onset).
(Ministry of Health, 2012)
The main purpose of this research study was to design a system to detect the plasma leakage in patients with DHF by analyzing patients’ medical records .Further, by using this system doctors can intervene early treatment of shock.
In recent years machine learning methods have been widely used in medical diagnosis. Medical diagnosis is one of major problem in medical application. Several research groups are working world wide on the development of neural networks in medical diagnosis. Neural networks are used to increase the accuracy and objectivity of medical diagnosis.Detecting plasma leakage is considered as a non-linear problem that shows the complex causal relationship between the variables. However, an artificial neural network that is suitable for problems of extreme complexity not addressable with conventional technologies, either by the conventional computer programming or statistical method.
In this research project multilayer feed forward neural network was used to train and test medical records of patients with DHF and DF. The trained network used to test more records of DHF patients to see the network performance and in order to make system practical to use in a real time hospital setting. The diagnostic performance of the proposed network is validated with Receiver Operating Characteristics (ROC) analysis to evaluate the sensitivity and specificity.
Currently, clinical identification of plasma leakage is difficult until DHF develops. The most common method of monitoring leakage relies on identification of haemo-concentration, determined by tracking changes in HCT measurements, with a rise of more than 20% from baseline considered evidence of significant leakage. However, this method can be rather insensitive, particularly if the patient is receiving parenteral fluid therapy, and it is also limited by the fact that an individual’s baseline value is rarely known.
(Ministry of Health, 2012)
Studies using ultrasound have demonstrated that pleural effusions, ascites and gall bladder wall oedema are common during the critical phase, and correlate with disease severity. In addition, serial ultrasound studies indicate that subclinical plasma leakage can be detected as days 2 to 3 of fever, and is better at predicting likely disease progression than other marker of plasma leakage such as HCT measurements. Gallbladder wall oedema appears to precede the development of ascites and effusions, and may therefore be a helpful early predictor of outcome. Thus ultrasonography is a useful monitoring tool, and where available, should be considered in the overall assessment during the febrile phase. However, there are certain limitations, particularly the lack of defined normal ranges for the parameters of interest, the variability in measurements obtained by different operators, and the lack of specificity of the findings.
(Srikiatkhachorn, Krautrachue, Ratanaprakarn, & al, 2007)
Chest X-ray is recommended to increase the sensitivity of detecting pleural effution. Pleural effusion detected clinically may not be obvious in a Chest X Ray (CXR)-PA, but may be seen only in a CXR right lateral decubitus film. (Ministry of Health, 2012)
An expert system can be divide into two sub-systems the inference engine and the knowledge base. The knowledge base represents facts and rules. The inference engine applies the rules to the known facts to deduce new facts. Inference engines can also include explanation and debugging capabilities
The pulse is how many times a minute that our arteries expand and contract in response to the heart. This rate is exactly equal to the heartbeat.
An artificial neural network is a mathematical model or computational model based on biological neural network. In other words, it is an emulation of biological neural system. An ANN is a network of highly interconnecting processing elements (neurons) operating in parallel.
Natural neurons receive signals through synapses located on the dendrites or membrane of the neuron. When the signals received are strong enough (surpass a certain threshold), the neuron is activated and emits a signal though the axon. This signal might be sent to another synapse, and might activate other neurons.(Gershenson,2003)
Artificial neuron is a highly abstracted model of the natural neuron. Inputs of artificial neuron behave like synapse of a biological neuron which are multiplied by weights (strength of the respective signals), and then computed by a mathematical function which is called Transfer function (also known as Activation function) in order to determine the activation of the neuron. The model of a neuron also includes an externally applied bias (threshold) that has the effect of lowering or increasing the net input of the activation function.
A MLF neural network consists of neurons that are ordered into layers. The first layer is called the input layer, the last layer is called the output layer, and the layers between are hidden layers. A neural network that has no hidden units is called a Perceptron. However, a perceptron can only represent linear functions, so it isn’t powerful enough for the kinds of applications. A multilayer feed forward neural network can represent a very broad set of nonlinear functions. Therefore, it is very useful in practice.
The behaviour of an ANN depends on both the weights and transfer function that is specified for the units. There are three transfer functions most commonly used for multilayer networks.
Supervised learning is an approach to find the input-output relationship based from the training using a set of data. Fig. 2.6 represents the block diagram of supervised learning. Learning system is fed with the input data and generates output, which is then compared with the target to compute the error signal by arbitrator. The error is sent to the learning system for further training until the minimum value of error is generated. (Muhammad Akmal Sapon, 2011)
The backpropagation algorithm is used in feed-forward ANNs. Artificial neurons are organized in layers and send their signals “forward”, and then the errors are propagated backwards. The network receives inputs by neurons in the input layer, and the output of the network is given by the neurons in the output layer. There may be one or more intermediate hidden layers. The backpropagation algorithm used for supervised learning. The network computes the error that is the difference between output and desired target and the backpropagation algorithm calculate how the error depends on the input, output and weights. The backpropagation technique reduces this error, until the ANN learns the training data.
In this research project Microsoft Excel 2010 was used to analysis the collected data and MATLAB R2013a (22.214.171.1244), 64-bit(win64) software was used as a tool to implement and to train the Neural Network.
The records of 10patients with DHF and 6 patients with DF from September 2013 to April 2014which are obtained from centre for clinical Management of Dengue and Dengue Hemorrhagic fever government hospital in Negombo. The data consists of total 1081 instances which 164 instances belonged to the leaking phase and 139 instances belonged to the non-leaking phase. Each data consists of 10 variables such as heart rate, systolic and diastolic blood pressures, PCV, Temperature and all are coded as numeric values. The patients are both male and female between 18 to 60 years old and who have over 50kg weight. These measurements are taken at equally spaced time points (hourly) since the patient was admitted to the hospital.
One of the most important parts in data preparation is to determine the best variables that contribute to the decision-making. The data selection step requires some detailed knowledge of the problem domain and the underlying data. Therefore, the selections of the variables are based on the advice of the doctors and also the review of the literatures. Even though there are quite a number of variables entered into the Observation Chart, only five variables are identified as the important variables that contribute to the detection of plasma leakage. They are as follows,
The training method was supervised training. Input vector contained 490 data for leakage phase and 591for non-leakage phase. The respective target for each was2-element class vector with a 1 in the position of the associated leaking or non-leaking. A two-layer feed-forward network with 20 sigmoid hidden neurons was created. The tansig(Tan-Sigmoid) is chosen as the transfer function for both hidden and output layers. The input vectors and target vectors are randomly divided into training, validation and test sets. From input vector 70% are used for training set,15% are used to validate that the network is generalizing and to stop training before overfitting, and the last 15% are used as a completely independent test of network generalization. The network was retrained until the network performance approach a satisfactory level (beyond 85%) of supervised training by using different training algorithms and increasing number of hidden neurons.
The trained neural network saved and it was used to test the new dataset. The new dataset consist of 50 leakage phase and 50 non-leakage phase data. Correct classifications and misclassifications were recorded.
Performance is measured in terms of mean squared error, and shown in log scale below figure 5.1. It rapidly decreased as the network was trained. Performance is shown for each of the training, validation and test sets. The version of the network that did best on the validation set is was after training.
This figure does not indicate any major problems with the training. The validation and test curves are very similar.If the test curve had increased significantly before the validation curve increased, then it is possible that some over fitting might have occurred.
Confusion matrix contains information about actual and predicted classifications done by a classification system for supervised learning system.
In confusion matrix, diagonal cells (in green cells) show the number of cases that were correctly classified, and the off-diagonal cells (in red cells) show the misclassified cases. The blue cell in the bottom right shows the total percent of correctly classified cases (in green) and the total percent of misclassified cases (in red). In this study, as shown in above figure 5.2, accuracy of training, testing and validation process are 90.9%, 80.9% and 79.0% respectively. Overall accuracy for correct classification is 87.6% and misclassification is 12.4%.Therefore the results show fairly good recognition.
The coloured lines (green and blue) in each axis represent the ROC curves for training, testing and validation. The ROC curve is another visualization of quality of the network. It is a plot of the true positive rate (sensitivity) versus the false positive rate (1-specificity) as the threshold is varied. A perfect test shows points in the upper-left corner, with 100% sensitivity and 100% specificity. In this study, the network performs fairly good.