Keywords

1 Introduction

Our health administrations are not that tremendously equipped with skills. The health monitoring framework is important to identify sicknesses effortlessly by just putting dataset of patients. This proposed framework can identify the diseases in patients with more accuracy when contrasted with past frameworks which took care of this issue. In this, health monitoring framework, for identifying the illness, has been led by particular indications related, for identifying the infections with more exactness. So, in this paper artificial neural networks (ANNs) used in health monitoring system. NN shows a seriously diverse way to deal with utilizing PCs as a part of the working environment. NN is utilized to study pattern and relationship in information.

1.1 General Theory of Artificial Neural Networks

Artificial neural networks are no more than an interconnection of artificial neurons. ANNs learn the relations between selected inputs and outputs from the previous experiences. ANNs also perform their tasks simultaneously (i.e., parallel processing), which makes ANNs very fast. A typical ANN can identify and learn the relationships between the inputs and the outputs of a nonlinear multidimensional system (see Fig. 1).

Fig. 1
figure 1

Nonlinear multidimensional system

Full size image

The neural network is an imitation of the neurons in the human nervous system. The human brain consists of over 100 billion neurons and over thousand times more synapses (the interconnections between the neurons). A tree-like structure extends from neuron cell that accepts inputs from other neurons. They are called dendrites. The carrier of the outputs from neuron to the dendrites in another is called an axon.

In a very simple form, an artificial neural network may be considered as a directed graph with nodes and weighted connections between the nodes. The network has one or more input nodes to accept sensory inputs and one or more output nodes to get back the response. The output from the network is a function of the input variables and the corresponding connection weights between them. Compared to the biological neuron system, the nodes are identical to the cell body, the connections are identical to the synapses, and the connection weights are identical to the synaptic efficiency.

The general classification of networks is done on the basis of how these connections are established and how the connection weights are modified to minimize some error measure on the response of the network to a stimulate. In some networks like the Hamming net or Hopfield net, the weights are fixed. However, in most other networks, the connection weights are updated in the training process. The process of minimizing the error measure by weight update is known as learning. Some networks have the ability to learn their own and are thus known as unsupervised networks.

On the other hand, supervised networks require a training set that is a set of possible stimulants and the corresponding output expected from the network. The later is thus a form of interactive learning in which, as a student learns from the teacher, the network adjusts its internal connection weights so as to produce the exact answer expected in the training set for each stimulates. The training in supervised networks is evaluated by another set of data known as a test set that is identical to the training set but contains stimulants not seen in the training process. It is possible to picture the training process as some curve fitting or parameter estimation process for the series expansion of a function. If the training has resulted in a proper convergence of the network on the training data, one should also expect a minimum error in the testing process. If that in the case, the network has a good generalization ability and the learning is ideal. The second possibility is that the network produces very good results on the training set but fails considerably on the test set. This is typical of what we call over fitting of data.

1.2 Artificial Neural Networks Training Methods

The learning procedure is called training and is performed by principles, which are grouped into the two fundamental after sorts:

  • Supervised learning: where both inputs and outputs are known, this implies the system can decide its prescient execution from given input.

  • Unsupervised learning: where the outputs are not known and the neurons, however, need to figure out how to oversee them, this technique is not considered in this.

1.3 Characteristics

Artificial neural networks (ANNs) or connectionist systems are computing systems inspired by biological neurons, performing likewise as a human brain henceforth, the trademark incorporates the ability for storing information and making it available for utilizing at whatever point essential, liking to recognize designs, even in noise, bent for taking past experiences and make inference and judgments about new circumstances.

1.4 Applications

  • Capacity to display straight and nondirect frameworks without the need to make presumptions certainly;

  • Data preparing, including filtering and clustering;

  • Computational neuroscience and neurohydrodynamics;

  • Forecasting and expectation;

  • Estimation and control.

2 Literature Review

Konstantina et al. in [1] proposed a prognosis and prediction of cancer using application of machine learning (ML). ML has a unique ability to detect the key feature from complicated datasets. Many techniques are preset to predict the disease like Bayesian networks, decision tree, and ANN. But according to this paper, ML method is more convenient to understand the cancer progression. ML also has a branch of artificial intelligence. Its aim is to design a model which can be used for prediction, estimation, and classification. For classification, author analyzed the model base on sensitivity, accuracy, and specificity. Validation method used is—fivefold cross-validation, cross-validation, and hold out. In this paper, author compares the work which all published recently for cancer prognosis by using ML methods. This comparative study will help us to opt for best method to design the system.

Giduthuri Sateesh et al. [2] presented metacognitive learning in a radial basis function network for classification problems. This system is inspired by metacognitive principles which have two components. One is metacognitive and second is cognitive component. To verify the result of a designed system, they choose low and high features small and large number of dataset. The dataset is from UCI machine learning. In this, author works on the dataset of liver disorder, breast cancer, and diabetes on low-dimensional feature which consist of around 200 samples this work shows and improvement of 7 and 14% as compared to previous result. The result from this sample is around 75%. In this, result used different–different method such as sequential multi-category RBF network, extreme learning machine (ELM), metacognitive RBF network (MCRBFN), self-adaptive resource allocation network (SRAN).

Amin et al. [3] in this paper evaluated that data mining is more appropriate technique for prediction of heart diseases. This technique is widely used for prediction and detection of many diseases, to get good accuracy. The important application of this diseases is to determine heart disease, based on the input attributes such as hypertension, age, obesity, tobacco, alcohol intake, and physical activity. This paper used two techniques for data mining, i.e., genetic algorithm and neural networks. This technique is implemented for diagnosis of 50 patients using MATLAB software. A multilayer feed forward network used, which consists of input, hidden, and output layer, first to initialize the weights and train LM function used to update weight and bias values. Author used 70% data for training and 15% for testing. The result which author got from this software is 89% accurate.

Ankeeta et al. [4] presented a neural network system for diagnosis of heart disease; heart diseases are classified in four ways—normal person, stroke one, stroke two, and stroke three. The author used multilayer feed forward network, this can be done in two steps, training and testing, the 13 input parameters are used in this to get the best output, and it repeats the training so many times to get the accuracy. Dataset taken from the Cleveland Data uses 297 training pairs. To map the input attributes value to the output target value, two-layer feed forward back propagation network is used, hidden layers are 15 and input layer 13. Tansig is used in training function. In the end, the work is classified in four classes—Class 0, normal person, Class 1: first stroke, Class 2: second stroke, and Class 3: third stroke.

Sonawane et al. [5] presented a method in the medical field for the diagnosis of heart-related problems using neural network. There are 13 features that are used in this paper, and the system is trained by back propagation algorithm. A large number of patients are increasing everyday because of heart diseases; this is the reason that he worked for this diseases. Back propagation network first initialized the weight of each neuron. Then, receive output signal from the training data and send it to the hidden unit. This hidden unit computes the value of net input by using the below equation.

$$Z = V_{oj} + \sum {\left( {i = 1\;{\text{to}}\;N} \right)X_{i} \cdot V_{ij} } .$$
(1)

Then, the output i is calculated by using the activation function. The dataset contains 303 data values out of which 164 are related to healthy category and 139 are related to heart diseases. From 303, 212 values are related to training and other 91 are used for testing. The machine gives 98% accuracy when 20 neurons are used for 1000 iterations.

Jasdeep et al. [6] approached different methods: scaled conjugate gradient and Levenberg-Marquardt propagation. These different methods are used to diagnose the thyroid diseases. The thyroid dataset is first trained using LM propagation and output is noted. Then, the dataset is trained by conjugate gradient back propagation. The aim of this paper is to represent the general point for the use of ANNs in medical field. This method is also used for classification, forecasting, and problem solving. This also has three layers. Hidden layer consists of 20 neurons. This paper consists of 7200 patient dataset. This dataset is divided into three categories—normal, hyperfunction, and subnormal functioning. For training, 5040 samples are used, 1030 are used for validation, and 1030 are used for testing.

Gokul et al. [7] presented the application of a fully complex-valued radial basis function network (McFCRBF) and extreme learning machine (ELM) for the diagnosis of Parkinson’s diseases. The two components of this are cognitive component and metacognitive component. It is a neurological degenerative effect. This disease slows down the movement, and it affects also speaking and writing. In this, result is based on the unified Parkinson’s disease rating scale (UPORS), and its values range from 0 to 176, in which 0 represents the healthy and firm condition and 176 represents the unhealthy or disability condition. The scale is based on three factors—mood, behavior, and daily leaving activities. The dataset contains 804 samples out of 575 are used for training and 229 are used testing. This contains four inputs and one target. Model is trained for 1000 epoch. From this paper, they conclude that MC-FCRBF is more appropriate method as compared to FC-RBF and ELM.

Ayush et al. [8] illustrated that the diabetes is worse than all other diseases. Diabetes causes many diseases such as blindness, Alzheimer, and kidney failure. This paper illustrates that human lifestyle such as sleeping habit, physical activities, eating habit plays a major role for any disease. Diabetes is a chronic disease which occur either when pancreas does not produce enough insulin or when the body cannot used it properly the insulin produced. The questionnaires set made for interaction with doctor, and then these questionnaires asked localities to collect data. It depends on amount of sugar intake, physical activities performed, fried food intake, rice intake and sleeping time, etc. These are the important factors which cause diabetes. Other factors are BMI (body mass index), hereditary diabetes history of chronic disease. Two types of dataset were prepared—diabetic person and nondiabetic person. K-fold cross-validation method is used for result validation. In that, k = 5 (means data divide into five parts). By this method, the author got the accuracy of 75% correct from the collected dataset.

Agarwal et al. [9] discussed the most effective and friendly model for expert to help in medical field. The model was motivated by biological nervous system. ANN can be categorized in many types—single-layer feed forward network, multilayer feed forward network. ANN has the ability of learning. ANN has three types of learning process—supervised, unsupervised, and reinforcement learning. This study illustrates the usefulness of artificial neural network technique in the diagnosis of cancer. In ANN, two methods are used for best result—MLP gives 97% accuracy, PNN gives 96% accuracy, and ART shows 92% accuracy.

Panduranga et al. [10] observed that data mining is more powerful tool to diagnose the diseases. They collected dataset from north coastal regions. Dataset of 504 people used and this dataset consists of 56 input attributes. Main work of data mining is to extract the pattern from dataset and convert it to gainful data. For this, MATLAB software is used to analyze the information. This method increases the accuracy of data. In this, probability neural network method is used to analyze the dataset from which confusion matrix is computed and trains that network to get the appropriate result. Results are in the form of receiver operating characteristic (ROC) plot, a plot of sensitivity versus specificity. The R value from this method is 0.69%, and the means accuracy is 69% (Table 1).

Table 1 Comparison of different techniques
Full size table

3 Methodology

The information analysis might be performed by neural systems. There is one neural system display utilized as a part of this examination: back propagation systems. It gives a computationally proficient technique for changing weight in a feed forward network, with differentiable activation work units [11]. The preparing of a system by back propagation includes three stages: the feed forward of the input training design, the calculation and back propagation of the related error, and adjustment of the weight. There are three noteworthy strides in the neural system: preprocessing, architecture, and post-processing.

  • In preprocessing, data are gathered that could be utilized as the input sources and output of neural systems [12,13,14]. This information is initially standardized or scaled for decreasing the variance and noise.

  • In design, an assortment of neural system models is fabricated that could be utilized to catch the connections between the information of sources of input and output.

  • In post-processing, distinctive systems are connected to the estimating results to boost the ability of the neural network prediction (Fig. 2).

    Fig. 2
    figure 2

    Design steps for ANN

    Full size image

The diagnosis of diseases is done by collecting data. Than data uploaded in MATLAB so it will pass from training and testing part. After its result if the result is more than a threshold value, then the practitioner assumed that person have chances of that diseases and recommends the patient to go for a laboratory test to assure about the result. By the test, it confirmed that which particular disease we have and what is the stage of that (Fig. 3).

Fig. 3
figure 3

Input and output attributes

Full size image

3.1 Simulation Tool Used MATLAB

The name MATLAB remains for matrix laboratory. In MATLAB, neural network toolbox is one of the usually utilized, effective, industrially accessible programming instruments for the advancement and outline of neural systems. The product is easy to understand and grants adaptability and accommodation in interfacing with different tool kits in a similar situation to build up a full application [15]. It can simply open by command nntool in MATLAB command window.

4 Result

Parameters are as follows:

  • Network Type = FeedforwordBackprop;

  • Train Function = TRAINLM;

  • Adaption Learning Function = LEARNGDM;

  • Performance Function = MSE;

  • Numbers of Layers = 2;

  • Hidden neurons = 100.

After, assigning the value of these parameters. Next, divides the samples in training, validation, and testing. Here, we divide into the ratio of 90, 5, and 5%, respectively.

Figure 4 shows a simulink diagram of neural network in which x1 shows its input and y1 indicates a output for this network and Fig. 5 shows a neural network training model in this first layer is input layer [16], 11 inputs are given to this system then, its feedback is given to the middle layer which is hidden layer neurons taken in this layer are 100. After this step, it will preprocess the information and give its result at the output layer. So in this model, we already set its output to 3.

Fig. 4
figure 4

Simulink diagram of NN

Full size image
Fig. 5
figure 5

Neural network model with different layers

Full size image

Then, training NN train tool plots the various state, i.e.,performance, training state, error histogram, and regression [17,18,19].

Here, one plot is attached which is neural network training regression plot.

Figure 6 represents training, validations, and testing graphs of regression analysis. In this, training shows an accuracy of 98.34%, validation shows an accuracy of 19.41%, testing shows an accuracy of 15.61%, and overall result for this implementation is 88.67%.

Fig. 6
figure 6

Regression plot

Full size image

5 Conclusion

This paper represents a tendency to access or utilization of ANN system for prediction of medical diseases. The idea of artificial neural systems in the field of medical is considered not extremely developed; the considerable measure of exploration is going on [20]. Thusly here, we designed simulated neural systems for diseases determination which prepared to utilize feed forward back propagation method. This model has been tried on a dataset that included patient data for determination of heart diseases, liver disorder, and diabetes datasets gathered from a UCI and nearby doctors. The parameters such as age, sex, blood pressure are important for prediction of disease in human health monitoring system. However, an addition of more feature such as HIV status, alkaline and phosphate added more carefulness to this system. From the results, it is observed that multilayer perceptron necessitated less number of neurons, less computation time as compared to other methods. ANN is adaptive, and this makes diagnose more reliable. In this system is trained first by using of datasets gathered form various sources, and then results are measured. Artificial neural systems illustrate noteworthy outcomes in managing and working with medical diseases dataset information. Results construed that this conclusion neural system-based model could be an important investigation of neural systems in the field of medical determination of diseases. In this architecture, the success rate of 88.67% was achieved.

6 Future Work

The proposed health monitoring system is based on MLP-NN that is more efficient. This system can be used in a wide range of medical field to help the professionals. But, the research is never ending process; a new beginning is always waiting. So, as a future work, more attributes can be used for training the system to increase the efficiency and accuracy of this system, which gives more appropriate results as compared to this. Also, new techniques can be applied for this.