M-GWO Algorithm to Predict Risk of Silent Heart Attack of Diabetes Patients - Cardidiabetes Model

As healthcare data is exponentially growing day-by-day in globe. Traditional model not sufficient to spread awareness in population for various new diseases worldwide. Therefore, diagnose critical diseases accurately was hard in the traditional health model. In case of emergency diagnose and provide treatment in minimum time to the patients reside at a remote location also tough for doctors. Traditional medical records were not measured accurately because of human error. In traditional health system diagnose chronic diseases efficiently was time taken process that may increase death ratio in OT. The key challenge is to handle the healthcare data in real time.

In traditional healthcare system, patients visit to hospital for their health issues. In critical cases, patients sometimes could not be able to reach hospitals. If any, healthcare system can access the data from remote location. It will help to save life of critical patients. Problem in traditional healthcare systems are as follows:

Real-time healthcare data is continuously changing, so old model cannot fit with new data. Thus, model has to be altered.

Volume of real-time data is too high, traditional tree algorithms cannot work as tree height and width would increase exponentially.

Another point of discussion is health issues of patients suffers from diabetes type 2. Diabetes type -2 patients have the high risk of silent heart attack and hyperglycemia. Which requires instant treatment. If such, patients can be diagnosed and treated from remote location could be helpful to save their lives. The possible approach is explored to overcome all problems in traditional model.  

Proposed model dinocarid consist of three phases:

GEETN (Pre-processing technique) is a novel approach,

M-GWO (Modified-Grey Wolf Optimization) important feature selection,

Stacking (Train and test Model). To accomplish the diagnosis of Diabetes + cardiac profile of patients both the model contributed successfully to produce efficient outcomes.

Figure 1 shows the proposed model where the process consists various steps. The process of filter the data is based on statistics and list is sorted towards the leaf node splitting of dataset is completed by arguments values.

5.1 Pre- processing technique (GEETN)

This proposed novel technique (GEETN) is to pre-process the raw data consisting of findings of cardio vascular and diabetes. First attributes are labelled based on reading of attributes - (Label 0: normal range of attributes, Label 1: borderline and above and Label 2: high risk) of cardiac arrest due to diabetes mellitus.  At first all the missing values were identified in raw healthcare data. Imputation is conducted with the standard values of attributes based on the standard healthcare defined by medical research center.

[ LDL = TC – HDL – 0.2*Triglycerides] is one of the of calculation is applied. Another example blood glucose level ranges below 80 considered as normal and label = 0, moderate range 80 from 180 label = 1, and high glucose above 180 label = 2. In similar manner all attributes have been calculated using (Eqs. (1)-(2)). Outcomes is labeled using Eq. (3).

Pre – processing – algorithm (GEETN) - The missing and null value of dataset has been changed on the basis of standard value of attributes defined by doctor. Second challenge is to process real time health data streaming. Figure 2 shows how to handle the null value is also matter of concern. Because any value in health data is specific health reading Why….. ? – Health attributes can be replaced with any random value and not with mean average value. It cant be 0 either.

1.png

Figure 1. Proposed model – flow chart representation of model

2.png

Figure 2. Categorical dataset

The following steps is applied to calculate GEETN.

(1) All value of attributes has been calculated for its normal range of health parameters.

(2) Health dataset is pre-processed using the normal range of each attribute. Then the outcomes are calculated according to the value of each attribute.

(3) The proposed method then identify as the null value in healthcare dataset and the healthcare dataset based on the novel approach.

(4) Imputation method for categorical columns is applied.

(5) Impute the missing value using GEETN.

(6) The missing value is first detected and replaced by calculating values based on other parameters.

(7) GEETN is applied on attributes those are correlated with other attributes. For example, cholesterol value predicted by blood strain is correlated with (LDL, HDL, VLDL). Likewise, Hb1ac.

(8) Imputation is evaluated on the basis of health value of different attributes.

(9) Dependent variable is calculated by using following equation:

(output)^ $n=\sum_{-}(i=0)^{\wedge} n$ (input_attributes)   (1)

(10) Further the output (dependent variable) is labelled according to the range of abnormality in diagnosis. Equation to label dependent variable:

TN= Total number of labels considered in range of abnormality

y= average of TN

z= $\sum_y^{T N} Z$

$\mid$ label $_{\text {outcome }}: x \lvert\,=\left\{\begin{array}{lr}x=0, & x=0 \\ x=1, & x>0, x \leq y \\ x=2, & x \geq z, x \leq z\end{array}\right.$      (2)

(11) Output is the new leaf node attribute from all observations of input attributes that has been defined for normal health of patients.

$(\text { output })^n=\sum_{i=0}^n($ input_attributes) For normal healthy readings of patient. （3）

5.2 Grey wolf optimization

Typically, grey wolves exhibit a preference for living in packs, with an average group size ranging from 5 to 12 individuals. Within the social structure of a grey wolf pack, there are stringent rules governing dominance and hierarchy. The pack consists of the following roles:

1. Alphas, serving as leaders, are responsible for decision-making, and their directives are followed by the entire pack.
2. Betas, subordinate to alphas, assist in decision and other group happenings. Betas can be either gender, and they often emerge as strong contenders for assuming the alpha role.
3. Omegas play the role of scapegoats and are required to update all dominant wolves. They also considered as last in line to partake in meals.
4. Alphas and betas may have authority over deltas, but deltas exert control over omegas. This group comprises scouts, sentinels, elders, hunters, and caretakers. Scouts vigilantly watch over territorial borders, warning the pack of potential threats. Sentinels are dedicated to safeguarding and ensuring the safety of the pack. Elders, having previously held alpha or beta roles, bring experience and wisdom. Hunters collaborate with alphas and betas in capturing prey to sustain the pack. Caretakers, on the other hand, attend to the needs of weak, ill, and injured pack members.

The framework of Grey Wolf Optimization model, the most optimal solution is denoted as alpha $(\alpha)$. The subsequent two best solutions are referred to as $(\beta)$ and ( $\delta$ ) delta [18]. The remaining solutions for another candidate are collectively designated as ( $\omega$ ) omega. Hunting [2] pattern is provided by these four entities- $\alpha, \beta, \delta$, and $\omega-$ each following one of the three aforementioned candidates [4]. To simulate the encircling pack's behavior necessary to hunt the prey, the mathematical modeling of this behavior is represented by the following equations, namely Eqs. (4)-(7).

$\vec{P}(t+1)=\vec{P} \rho(t)+\vec{W} \cdot \vec{Z}$  (4)

In Eq. (5), $\vec{D}$ is represented with iteration number $\mathrm{t}, \vec{W}$ and $\vec{Y}$ are vectors of coefficients. $\vec{P} \rho$ denotes the prey position, while $\vec{P}$ signifies the grey wolf position [6].

$\vec{W}=|\vec{Y} \cdot \vec{P} \rho(t)-\vec{P}(t)|$     （5）

The $\vec{W} . \vec{Y}$ vector are calculated as in Eqs. (6) and (7)

$\vec{x}=2 x \cdot \overrightarrow{r 1}-x$     （6）

$\vec{Y}=2 \overrightarrow{r 2}$ （7）

The process involves linearly decreasing the parameter 'a' range 2 to 0 throughout iterations, with 'r1' - 'r2' [18] representing random vectors within the range [0, 1]. Generally, the alpha guides the hunting, with occasional participation from beta and delta. Wolves' positions are updated by expression as in Eq. (8).

$\vec{P}(t+1)=\frac{\overrightarrow{P_1}+\overrightarrow{P_2}+\overrightarrow{P_3}}{3}$       (8)

$\overrightarrow{P_1}, \overrightarrow{P_2}, \overrightarrow{P_3}$ are define as in Eqs. (9)-(11), respectively.

$\overrightarrow{P_1}=\left|\overrightarrow{P_\alpha}-\overrightarrow{W_1} \cdot \overrightarrow{Z_1}\right|$     (9)

$\overrightarrow{P_2}=\left|\overrightarrow{P_\beta}-\overrightarrow{W_2} \cdot \overrightarrow{Z_1}\right|$  (10)

$\overrightarrow{P_3}=\cdot\left|\overrightarrow{P_\gamma}-\overrightarrow{W_3} \cdot \overrightarrow{Z_1}\right|$  (11)

where $\overrightarrow{P_\alpha}, \overrightarrow{P_\beta}, \overrightarrow{P_\gamma}$ best 3 solution found in first phase with t iteration of swarm, $\overrightarrow{W_1}, \overrightarrow{W_2}, \overrightarrow{W_3}$ are mentioned in Eq. (3), and $\overrightarrow{Z_\alpha}, \overrightarrow{Z_\beta}, \overrightarrow{Z_\delta}$ Eqs. (12)—(14) defined, respectively.

$\vec{Z}_\alpha=\left|\vec{Y}_1 \cdot \overrightarrow{P_\alpha}-\vec{P}\right|$  (12)

$\vec{Z}_\beta=\left|\overrightarrow{Y_2} \cdot \overrightarrow{P_\beta}-\vec{P}\right|$  (13)

$\vec{Z}_\gamma=\left|\overrightarrow{Y_3} \cdot \overrightarrow{P_\gamma}-\vec{P}\right|$  (14)

where $\vec{Y}, \overrightarrow{Y_2}, \overrightarrow{Y_3}$ are define as in Eq. (5).

A concluding note regarding the GWO pertains to the adjustment of the parameter 'x'. The parameter 'a' undergoes a linear update during each iteration, transitioning from 2 to 0 in accordance with Eq. (15).

$\alpha=2-t \frac{2}{\text { Maxiter }}$  (15)

Modified Grey Wolf Optimization (M-GWO):

Wolf locations are continually adjusted to any point in space using the GWO technique. Certain unique issues, such feature selection, have solutions that can only be found in the ranges from {0,1} binary digits, by which a unique version of the GWO is needed. This proposed model proposes novel M-GWO for the job of feature selection. Each wolf is drawn to the top three best solutions by the wolves updating equation [2], which is a function of three location vectors, $\mathrm{X} \alpha$, xβ, and $\mathrm{x} \delta$. The pool of solutions in the M-GWO is always in binary form, with every solution located on a hypercube's corner.

$P_i^{t+1}=\operatorname{Crossover}(P 1, P 2, P 3)$   (16)

where Crossover [19] considered as preferred cross over among solutions - x; y; z and P1; P2; P3 vectors of binary, indicating impact on movement of wolf towards alpha, delta, and beta grey_wolves in sequence, are represented by P1, P3, and P2. The calculations for these vectors are determined through Eqs. (17), (20), and (23), respectively.

$P_i^d=\left\{\begin{array}{l}1 \text { if }\left(P_\alpha^d+\text { bstep }\right. \\ 0 \text { Otherwise }\end{array}\right.$   (17)

Here, $P_\alpha^d$ denotes the alpha wolf’s position vectors in dimension d, and $b s t e p_\alpha^d$ is represented as binary step for d dimension, which can be computed according to the formula presented in Eq. (18).

bstep$_\alpha^d=\left\{\begin{array}{l}1\text { if cstep}_\alpha^d \geq \text { rand } \\ 0 \text { Otherwise }\end{array}\right.$    (18)

In this context, "rand" represents a randomly generated number from a uniform distribution. Meanwhile, $\operatorname{cstep}_\alpha^d$ stands for the continuous-valued step size [20] associated with d dimension. This step size may be determined by employing a sigmoidal function [2], as outlined in Eq. (19).

${cstep}_\alpha^d=\frac{1}{1+e^{-10\left(A_1^d\left(D_\alpha^d-0.5\right)\right.}}$  (19)

where $W_1^d$ are evaluated using Eqs. (3), and (9) in the $d$ dimension.

$P_2^d=\left\{\begin{array}{l}1 \text { if } x_\beta^d \geq 1 \\ 0 \text { Otherwise }\end{array}\right.$      (20)

In this scenario, $P_2^d$ signifies the beta wolf's position vectors of wolf in d dimension, and a binary $bstep_\alpha^d$ step in mentioned in d dimension, the computation is specified in Eq. (21).

 $bstep_\beta^d=\left\{\begin{array}{l}1 \text { if cstep}_\beta^d \geq \text { rand } \\ 0 \text { Otherwise }\end{array}\right.$  (21)

In this context, "rand" denotes a randomly generated number from a uniform distribution. Additionally, $\operatorname{cstep}_\alpha^d$ represents the continuous-valued step size [2] associated with dimension d. The calculation of this step size involves the utilization of a sigmoidal_function, as described in Eq. (22).

$\operatorname{cstep}_\beta^d=\frac{1}{1+e^{-10\left(A_1^d\left(D_{\beta^{-0.5}}^d\right)\right.}}$    (22)

where W1, Zd, and d for β are calculated using Eqs. (3), (9) in the (d) dimension.

$P_3^d=\left\{\begin{array}{l}1 \text { if }\left(P_\delta^d+bstep_\delta^d\right) \geq 1 \\ 0 \text { Otherwise }\end{array}\right.$ (23)

In this context, $P_\delta^d$ designates the delta wolf's position vectors in dimension (d), while $b s t e p_\beta^d$ represents a binary step for d dimension. The computation of this binary step is outlined in Eq. (24).

$bstep_\delta^d=\left\{\begin{array}{l}1 \text { if cstep}_\delta^d \geq \text { rand } \\ 0 \text { Otherwise }\end{array}\right.$       (24)

The modified version is executed using two distinct methods. In the initial method, steps followed for top three best solutions are labeled and binarized. A stochastic crossover is then employed among these 3 fundamental moves for determining the grey wolf's next expected position. The next approach involves using a sigmoidal function to transform the continuously updated position. The resulting values are stochastically thresholder to ascertain the updated position of the grey wolf. Both methods for the modified Grey Wolf Optimization (M-GWO) are applied within the field of feature selection, with the objective of identifying a subset of all features that maximizes classification accuracy while minimizing the selected features. Figure 3 presents the proposed stacked Model.

3.png

Figure 3. Proposed stacked model

KNN:

KNN is acknowledged as forthright and logical machine learning technique. Nearest data points are compared. New data points are assigned to nearest neighbours. Where "k" is total neighbours needs to considered for calculation of data points. KNN is simple to use and understand. The KNN [21] is beneficial for small size of dataset, while apply is on bigdata may be disappointing.

Random Forest: RM is a well-organized Machine Learning technique. It is applied for both problems – classification and regression.  Tree is generated from randomly selected training data; group of decision tree are produced from dataset. afterwards, final prediction is produced by complete results of several decision tree.

Decision Tree (DT): An efficient ML approach based on numerous input features, model is constructed by selected tree and potential outcomes. Best suitable tree is selected among all iteration. In case of overfitting DT is complicated. 

SVM (Support vector machines): The aim of SVM is technique in which margin and hyperplane is used to classify dataset. Margin is used to decide boundary by taking support-vectors for hyperplane.

To detect the healthcare real time data (Based on Diabetes and cardiac diseases), algorithm applied on dataset those are best for feature prediction and splitting the dataset to predict the disease accurately. Comparison of traditional algorithm and proposed model applied for prediction of health disease. various algorithm such as KNN SVM_RBF, DT, RF, MLP [23] and proposed model. accuracy of 99 % MCC (70.35%), and f1 score (99.16 %). Overall performance of proposed model is outshined than as compared to traditional algorithms. features to evaluate risk of cardiac arrest due to diabetes. The graphical representation of all parameters shown in Figure 11 is well defined. Along with the performance, the proposed method also categories the number of patients according to their health status (Normal, Moderate and High risk) using proposed methodology. Novel approach applied in proposed work counts the health status of patients by checking the risk factor evaluated by attributes of diabetes and cardiac symptoms reading. proposed model algorithm gives 99% accuracy helps to predict the health disease at early stage based on patient’s data and precautionary measures may be suggested by doctors. The cardiodibet is based on prediction of cardiac + diabetes symptoms the classification of patient’s health status can be improved. Framework is successfully predicting cardiac arrest and the impact of diabetes over cardiac arrest. In future, framework can be able to predict the effect of diabetes on other vital organ those are not addressed in current studies like – kidney, eye retina etc. and predict the risk of damage in another vital organ.