Fuzzy Inference with Enhanced Convolutional Neural Network Based Classification Framework for Predicting Heart Attack Using Sensor Data

A heart attack is a major and common medical problem in the human being because of the various reasons and changed life style. A blood clot generally causes a heart attack by obstructing the heart's blood supply. Without blood, tissue dehydrates and deteriorates. A tightness or soreness in the chest, neck, back, or arms, as well as weariness, dizziness, an irregular heartbeat, and anxiety, are other symptoms. Atypical symptoms are more common in women than in males. Treatment options include anything from medication, stents, and bypass surgery to dietary adjustments and cardiac rehabilitation. The World Health Organization states that the leading cause of death globally is cardiovascular disease (CVD). In 2015, about 17 million people died as a result of CVD, accounting for almost 30% of all fatalities worldwide.

Several methods based on clustering, association rules [1], common patterns [2], sequential patterns [2], and classification were effectively applied to sensor data. The design and implementation of sensor networks, however, poses unique research issues because of their massive size (up to thousands of sensor nodes), haphazard and risky deployment, loss communication environment, constrained power supply, and high failure rate. These issues make using traditional mining methods impossible. Because mining has typically been centralized, expensive in terms of processing, and focused on transactional data stored on the disc, these issues make traditional mining approaches inapplicable. Microelectronic devices and wireless communication have advanced as a result of the installation of large-scale sensor networks and low-power sensors.

The most common of these conditions, coronary heart disease, resulted in nearly 360,000 heart attack-related deaths in the United States in 2015. The Centers for Disease Control and Prevention (CDC) estimate that an American has a heart attack every 40 seconds or so [3]. Additionally, the American Cardiac Association predicts that by 2030, the cost of treating heart disease will have doubled. In various application areas, including habitat monitoring [4], object tracking [5], environmental monitoring, military, disaster management, as well as smart environments, sensor networks have attracted significant interest due to their capabilities of ubiquitous surveillance.

Data mining techniques have been shown to be an effective tool for raising the effectiveness and quality of services (QoS) provided by WSN. In WSN, knowledge discovery has also been used to gather data about the immediate surroundings. This data is derived from sensor node data and behavioral patterns, which are produced from meta-data characterizing sensor behaviors.

Heart disease was the primary cause of almost 7 million of these fatalities, and industrialized nations accounted for more than 75% of all CVD deaths. Over 50% of fatalities from heart disease are male. Over 630,000 Americans each year die from heart disease in the United States alone, which accounts for 25% of fatalities. Congenital, coronary, rheumatic, and other types of heart disease are all included under the general term "heart disease." This study tells the important requirement for the automatic mechanism to predict CVD worldwide.

This research paper is arranged as follows: defined the literature work in Section 2. The proposed method is discussed in section 3. The experimental outcomes are discussed and evaluated in Section 4. Finally, the paper ends with Section 5.

This section provides the details of the classification and prediction frameworks for the diagnosis of heart attack disease based on the deep neural network architecture. Here, designing a modified convolutional neuro fuzzy inference framework for heart attack prediction. Proposed Neuro-fuzzy data processing system contains eight layers including two activation layers with activations functions i.e., ReLu and Softmax.

Each node of the input layer corresponds to one input variable, and since the first layer simply transfers input values to the following layer, it utilizes a linear function. Each input variable space is fuzzy quantized in the second layer. Second-layer neuron serves as a membership function and represents a term of an input linguistic variable. Although have other membership functions, such as trapezoidal, triangle, etc., we utilize the Gaussian function in this layer to increase the correctness of our design. Rule nodes in the third layer undergo learning using convolutional networks and activation functions. The fuzzy quantization and variable pooling for down sampling are represented by the fourth layer of neurons. The fifth layer is a fully connected layer with all filtered fuzzy rules. The sixth layer represents defuzzification of the rules which are filtered and finalized. The seventh layer represents the activation layer to convert the output into crisp values should fall in 1 and 0. The last layer i.e., eighth layer represents the output of the proposed framework.

The framework of the proposed research work is designed based on the conventional CNN, with a few convolutional layers, a fully connected layer, and fuzzy inference model. Every convolutional layer normally comprises of a convolution phase, a nonlinearity phase, and ultimately a pooling level. The systematic three-stage structure of the designed framework. During the first step of this, the input data is managed and fed into the fuzzification layer to render the fuzzy logic interpretation. The fuzzy convolution layer is subsequently convoluted with fuzzy logic interpretation that includes fuzzified convolutional layers. A min-max composition of fuzzified kernels with local fuzzified input characteristics is the Fuzzy convolution. From its input spatial features, each min-max composition derives higher-level fuzzy characteristics. The proposed framework showing how fuzzy convolution filters help to extract high-level characteristics from fuzzified input data. In the fuzzy domain, via the application of the fuzzy convolution series, the fuzzy convolution stage output feature sets are highly abstract attributes.

As some of the neurons should not get activated sometimes this RELU activation function keeps null for all negative values keeping them inactive when it is unnecessary. ReLu provides us this benefit. The activation function is given in (1)

$f(s)=\max (s, 0)$                 (1)

Imagine a network with initialised (or standardized) random weights and almost 50 percent of network activation due to the ReLu function (output 0 for negative values of x). This means that fewer neurons are firing (sparse activation) and there is a lighter network. The Softmax function is shown below mathematically in Eq. (2), where z is the input vector to the output layer. And again, j indexes the units of production, so j = 1, 2,…, K.

$\sigma(z)_j=\frac{e^{Z j}}{\sum_{k=1}^K \,\,e^{Z k}}$                   (2)

This function allows in determining the probability of more than one class at once. It is commonly used in the neural network as in the O/P layer and should sum up to 1. A neural network here is predicting whether the patient is suffering from disease or not. It will produce a probability, but it would do individually for each and every input. This function layer allows us to run many classes at once. The number of Softmax function nodes will be the same as O/P layer nodes. Here in the proposed model, build the output layer for predicting 0 or 1.

The proposed model takes the input of S(T) training samples and corresponding L(Y) labels, hyper-parameters αfc, αca, αcb, and αcw are the learning rates; N(B) batch number; dropout rate; I(N) training iteration number; C(L) number of convolution layers. And produces output of trained parameters and the model is designed with initializing the weight W and membership functions cx, cw, cy randomly. In the proposed model, few layers including input layer perform linear transformation on the input data S(T). Linear transformation Z is done according to the Eq. (3)

$\mathrm{Z}=w_T \cdot S(T)+b$                  (3)

Here in Eq. (3), w_T  Represents the weight, S(T) is the input data, and b is the bias value.

If the input matrix is A, several linguistic labels are given for every variable in A in association by membership functions. The fuzzy membership feature determines the degree that determines an input node membership to a specific fuzzy set. In Eq. (4), the fuzzy sets A are obtained via fuzzy relation of Eq. (5) which is computed by the max product procedure. It is linked to the probability that input and output data fit into the predetermined reference fuzzy records of M_pq.

$\mathrm{A}=$ fuzzification $\left(\mathrm{a}_{\mathrm{pq}} \mid \mathrm{c} \mathrm{a}_{\mathrm{pq}}\right)$                     (4)

Here p, q, in input matrix A represents indices of element a.

$\mathrm{a}_{\mathrm{pq}}=\text { possibility } \left(\mathrm{a}_{\mathrm{pq}} \mid \mathrm{M}_{\mathrm{pq}}\right) \max _{\mathrm{a \in A}}\left(M_{\mathrm{pq}} \delta\left(\mathrm{a}-\mathrm{a}_{\mathrm{pq}}\right)\right)$                    (5)

Here in Eq. (5), δ(a-a_pq) represents the Kronecker delta function.

Three processing stages are included in each fuzzy convolution layer: the fuzzy convolution stage, the nonlinearity stage, and the pooling stage. As given in Eq. (6), The fuzzy convolutional step occurs a method of employing fuzzy convolutional filters to fuzzy 2D information in that F_μ fuzzy convolutional filters are measured using formula (7).

$a_{\mathrm{pq}}=\sum_{p=0}^{i-1} \sum_{q=0}^{i-1} \mathrm{~F}_{\mu \mathrm{a}}(\mathrm{a}+\mathrm{p})(\mathrm{b}+\mathrm{q})$                    (6)

$F_{\mu a}=$ fuzzification(F)                  (7)

The Eq. (8) is a non-linear conversion of the acquired output from the fuzzy convolution point. The final step is a process called pooling, that is a brief statistic of immediate results with the function's extraction point. This phase enables the interpretation to be invariant to the conversion of the input, and in the meantime, the amount of the input to the subsequent fuzzy convolutional layer could be decreased. Crisp values are generated by the defuzzification layer from the characteristics extracted through pooling. Eventually, as an output classifier of the proposed framework, produces a completely linked layer feature.

The f(a_pq) represents the input activation function for the input to transfer to the next layer. This activation function activates the input and send some filtered output to the next fuzzy layer using the function in Eq. (8).

$f\left(a_{\mathrm{pq}}\right)=\max \left(a_{\mathrm{pq}}, 0\right)$                    (8)

$b_{\mathrm{pq}}=\sigma\left(a_{\mathrm{pq}}\right)$                     (9)

where, σ is the convolution layer's activation feature shown in Eq. (9). The completely connected layer of the proposed framework serves as a classifier through the crisp value gained in Eq. (10) being the input characteristics. With the center of gravity method from the defuzzification process. Here b_p  is the base of defuzzification's membership function, b_p^'  is the yield of the classifier given in Eq. (11).

$\operatorname{defuzz}\left(a_{\mathrm{p}}\right)=\frac{\sum c_b a_p}{\sum a_p}$                 (10)

$b_p^{\prime}=$ fullyconnected $\left(b_p\right)$                  (11)

The loss function used for measuring the output error is cross entropy as given in Eq. (12) here b_p  is the target b_p^' is the outcome of the classifier, and the sample count is S.

$\mathrm{G}=-\frac{1}{S} \sum_{I=1}^S\left[b_p \log \left(b_p^{\prime}\right)+\left(1-b_p\right) \log \left(1-b_p^{\prime}\right)\right]$                      (12)

By a conventional back propagation learning algorithm, the model parameters are trained with the cross-entropy loss function. The W_fc  updated as shown in Eq. (13) of the fully connected layer weight with an αfc learning rate.

$W_{f c}(\mathrm{k}+1)=W_{f c}(\mathrm{k})-\alpha \mathrm{fc} \,\,\partial \mathrm{E} / \partial W_{f c}$                   (13)

The c_b(k) centers for the membership functions of defuzzification are revised using Eq. (14), here pc_b  represents the updating center learning rate, $c_{k+1}$ and $\left(c_{k+1}\right)^{\prime}$ are the outcome target and the definite output of the model, accordingly.

$c_b(\mathrm{k}+1)=c_b(\mathrm{k})+\mathrm{p} c_b \nabla c_b$                    (14)

The center value c_w(k) and σ variance of the weight of the fuzzification membership function of the convolution layer is calculated by Eqns. (15)-(18) with a learning rate of $\alpha c_w$.

$c_w(\mathrm{k}+1)=c_w(\mathrm{k})+\alpha c_w \nabla \mathrm{W}_\mu$                   (15)

$\sigma_{c w}(\mathrm{k}+1)=\sigma_{c w}(\mathrm{k})+\sigma_{c w} \nabla \mathrm{W} \mu$                    (16)

where $\delta_{\mathrm{k}}=\left(\mathrm{W}_{\mu \mathrm{k}}\right)^T \delta_{\mathrm{k}} \mathrm{f}^{\prime}\left(\mathrm{i}_{\mathrm{k}}\right)$                        (17)

$\nabla W_{\mu \mathrm{k}}=\sum b_{p q} * \operatorname{rot} 180\left(W_{\mu \mathrm{k}}+1\right)$                     (18)

Here, f'(i_k) is the derivative function of i_k and rot180($W_{\mu \mathrm{k}}+1$) is the operator that rotates 180 degrees with respect to $W_{\mu \mathrm{k}}+1$. Eqns. (19) and (20) are used to change the mean and variance of the affiliation function of the fuzzification sheet and αcx is the rate of learning.

$c_{\mathrm{p}}(\mathrm{k}+1)=c_{\mathrm{p}}(\mathrm{k})+\alpha c_{\mathrm{p}} \nabla c_{\mathrm{p}}$                      (19)

$\sigma c_{\mathrm{p}}(\mathrm{k}+1)=\sigma c_{\mathrm{p}}(\mathrm{k})+\alpha c_{\mathrm{p}} \nabla c_{\mathrm{p}}$                       (20)

The training process for the proposed framework with a back-propagation strategy is summed up in above process. The good parameters of the proposed framework for the provided training dataset through a function list and target to obtain, employ small set of training procedures. The hyperparameter are like batch size, dropout rate, and learning rates are selected through empirical observation.