Classification of a New-Born Infant’s Jaundice Symptoms Using a Binary Spring Search Algorithm with Machine Learning

Most occurrences of jaundice in neonates are caused by the disintegration of cells, which release hematoidin into the blood, and the inability of the liver to adequately metabolise the bilirubin and make it available for elimination in the urine. High levels of hematoidin cause a xanthous staining of the whites of a new-born’s skin; symptoms include drowsiness and difficulty eating; consequences include convulsions, cerebral palsy, and kernicterus [1]. Pathologic jaundice is diagnosed in babies who are on their first day of life and have been present for more than two weeks, or if the newborn appears ill. It is common for neonates to acquire jaundice due to two causes [2]. To begin, the liver's hence foetal haemoglobin is degraded and replaced with adult haemoglobin, and bilirubin is eliminated as quickly as in an adult. Jaundice is caused by an excess of bilirubin in the blood, a condition known medically as hyperbilirubinemia. Second, if the neonatal jaundice does not improve with simple causes of liver disease in children, other options should be considered. These conditions include biliary atresia, Progressive familial deficiency, and others. Persistent neonatal jaundice is a medical emergency that must be treated right away [3].

Jaundice is caused by hyperbilirubinemia, which occurs in about 60% of babies and nearly 100% of preterm infants [4]. Jaundice is a somewhat common but usually harmless illness in neonates. Newborns aren't in danger of developing kernicterus unless the condition is misdiagnosed or treated late [5]. A chronic form of bilirubin encephalopathy known as kernicterus arises when bilirubin accumulates in the brain, causing irreparable damage [6]. Newborns at risk for developing must be correctly identified so that they can receive therapy as soon as possible. Therefore, protecting the baby from potentially harmful bilirubin levels is a top priority for doctors [7]. The risk of neonatal jaundice is currently evaluated using age-specific monograms that also factor in serum or other risk variables [8]. Studies have shown that there is a growing comeback, despite the fact that multiple methods are used to estimate the risk of developing brand-new hyperbilirubinemia [9].

Early diagnosis of jaundice in neonates and the related reduction in complications can be aided by knowledge of modifiable risk factors and the severity of the disease. There has been a rise in the number of babies being sent home from the hospital within the first 48 hours of life [10], and with such a brief postnatal hospital stay, jaundice may not have had time to manifest. Since many babies in Nigeria arrive at hospitals late with kernicterus, the country's neonatal mortality rate was 48 per 1,000 live births; an estimated 700 new-born’s died daily, totaling 284,000 deaths in a single year [11]. Neonatal hyperbilirubinemia is preventable, but only with prompt medical attention [12]. Parents should be able to spot neonatal bitterness and seek medical assistance for it immediately so that they can be discharged from the hospital sooner.

Data mining is one of the newest areas of science. It uses a wide range of statistical methods, ways to store data, AI tools, and pattern recognition, which is a branch of machine learning. Data mining, the application of statistical methods and AI to database management, is based on the ability to discover patterns and relations within massive datasets, which allows for the creation of models capable of assigning the class label to previously unlabeled cases [13]. Thus, data mining methods have been successfully implemented in many categorization problems [14]. Data mining's ability to unearth previously unsuspected connections has the potential to revolutionise our understanding of many medical conditions and motivate new studies in a wide range of scientific subfields. Data mining has helped the medical field thanks to the extremely high accuracy of the produced models [15]. Data mining has been shown to significantly improve the consequences achieved with existing approaches and contribute to new discoveries in several fields of medicine [16].

With the rise of big data and the advancement of computational control, machine learning techniques have emerged as a driving force in medical data analytics and knowledge discovery [17]. The need to apply machine learning algorithms to Big Data in order to uncover useful patterns has grown as the volume of patient data kept in healthcare databases has increased. Based on knowledge of preventable risk issues, this study employs machine learning to create a perfect organisation for the degree of jaundice in neonates. So, this paper focuses on creating a predictive model for the primary diagnosis of bitterness in newborns.

The contribution of the paper is given below:

• This research suggests using the XGBoost algorithm for precise jaundice illness forecasting. The BSSA model is used to choose the most useful features.
• The effectiveness of the suggested XGBoost algorithm is evaluated in relation to that of more standard machine learning approaches. The XGBoost algorithm is shown to be 3% more accurate than the current method.

The paper is organised as shadows: The associated work is described in Section 2. In the third section, we introduce the XGBoost algorithm we propose. In Section 4, we compare and contrast the proposed model to some commonly used validation methods. The study's conclusion and directions for future investigation are presented in Section 5.

3.1 Skin detection

Colour space, model, and classification technique all affect how well skin detection works. Skin detection is a critical but time-consuming task that necessitates careful consideration of colour spaces before proceeding. These studies employ a wide variety of colour spaces, from the familiar RGB model to the more experimental YUV and YCbCr colour spaces [25, 26]. The RGB colour model is widely used to represent colour images in a variety of settings. For real-time skin identification, YUV or YCbCr colour spaces are frequently utilised, and the transition from YUV to RGB is linear. Y stands for the luma portion of the hue. Luma is a measurement of how brilliant a colour is. This is a reference to how bright the colour is. This part is designed to work better with the human visual system. Chroma has a blue component (Cb) and a red component (Cr). When compared to green," Cb is the blue component," while "Cr is the red component," as the definition puts it. The human visual system has lower sensitivity to these factors. In this dataset, we provide images of new-born infants with and without jaundice.

Through the process of this study, we transform the infant's input image from RGB to grayscale and then from grayscale to binary. Images with the skin's surface modified have the relevant parts cut out using a region of interest (ROI). After the piece was cut out, the hybrid median filter was used to get rid of any background noise that was still there. To make a classification method for jaundice, you have to build mathematical proofs and cut down on the amount of data used to classify an input signal. These mathematical models of motion can be thought of as a reduction of the space occupied by a multidimensional space into which various input signals have been mapped. This process of reducing the number of dimensions is known as "feature extraction." Finally, the extracted feature set should only contain the most crucial information from the original signal.

3.2 Pre-processing median

The median of the candidate region is the grey level value found exactly in the middle of the ordered list of grey level values.

$S D=\sqrt{\frac{1}{N-1} \sum_{i=1}^N\left(x_i-\bar{x}\right)^2}$     (1)

Images need to be denoised before being processed since the presence of sounds reduces the detecting system's effectiveness. Here, we denoise the raw infant image using the hybrid median filtering algorithm.

3.3 Feature selection using BSSA

Discrete-time optimisation of any issue is possible with the spring search algorithm, which is based on the spring force laws. The spring force law is used to make it easier for people involved in the optimisation process to talk to each other. In general, the suggested optimisation approach can be applied to any optimisation problem where the solution exists. The objective functions are used to calculate the spring stiffness coefficients.

The spring force law, often known as Hooke's law, is represented by Eq. (2). Within a specific allowance for length change, nearly all springs behave according to this law.

$F_s=-k x$     (2)

where, k is the constant of the spring force, x is the degree of expansion or reduction, and F s is the force exerted by the spring. The two stages of SSA are as follows: Step 1: Establishing a simulated system in the issue area that uses discrete time, setting initial positions for objects, enforcing rules, and fine-tuning parameters; Step 2: Performing discrete-time iterations until termination requirements are fulfilled.

3.3.1 System development, implementing laws and tuning strictures

The initial step is establishing the system's space. Each point in the search space, a multi-dimensional coordinate system, represents a possible answer to the problem at hand. Each object in a set of searching operators is connected to every other object in the set via springs. In the search space, each object has a location, and the springs attached to it have stiffness coefficients. Each spring's stiffness coefficient is adjusted in accordance with the fitness values of the two nodes. In SSA, the laws of motion and the spring force law regulate the system.

Reflect a system which contains of m objects; the site of each object represents a point in the space and a solution to the tricky. The subsequent phrase displays the location of the ith object:

$X_i=\left(x_i^1, \ldots, x_i^d, \ldots, x_i^n\right)$ for $i=1,2, \ldots, m$     (3)

where, x id is the location along the ith item's dth dimension. All objects are initially scattered around the search area in the problem. Depending on the forces applied to each other by the springs, these objects are gravitating towards equilibrium (the optimal solution). The spring stiffness coefficients are calculated using the following equation.

$K_{i, j}=K_{\max }\left|F_n^i-F_n^i\right| \max \left(F_n^i, F_n^j\right)$     (4)

where, Fni and Fnj are the normalised objective functions of the objects I and j, K (i,j) is the stiffness coefficient of the spring positioned among objects I and j, K max is the extreme value of stiffness springs and stated based on the optimization problem. Following are some expressions that are used to normalise the objective functions:

$F_n^{\prime i}=\frac{f_{o b j}^i}{\min \left(f_{o b j}\right)}$     (5)

$F_n^i=\frac{\min \left(F_n^{\prime i}\right)}{F_n^{\prime i}}$      (6)

where, f obj is the value of the impartial function for the object iteration I and f obji is the value of the objective function for the object iteration i+1. If we define an axis for each dimension of an m-dimensional optimization problem, we can find out how each variable is projected onto each axis. By comparing, the stable points on each axis are located on the left and right sides of the object. Stable points for a given object are simply locations where it has a higher fitness relative to other objects. So, each object experiences a sum of two forces along each axis. The formulations for the forces acting to the right and left of object j in dimension d are as follows:

$F_{\text {total }_R}^{j, d}=\sum_{i=1}^{n_R^d} K_{i, j} x_{i, j}^d$      (7)

$F_{\text {total }_L}^{j, d}=\sum_{l=1}^{n_L^d} K_{l, j} x_{i, j}^d$     (8)

where, K (i,j) and K (i,j) are the initial and final velocities, respectively, and F (total R)(j,d) and F (total L)(j,d) are the resulting forces on the right and left sides of object j in dimension d, the number of stable points on sides of object j in dimension d, and the distances between object j and the stable points on sides of object j in (l,j) The following are the outcomes of applying Hooke's law in dimension d.

$d X_R^{j, d}=\frac{F_{\text {total }\,_R}^{j, d}}{K_{\text {equal }\,_R}^j}$      (9)

$d X_L^{j, d}=\frac{F_{\text {total }\,_L}^{j, d}}{K_{\text {equal }\,_L}^j}$      (10)

where, rightward translation (dX R(j,d)) and leftward translation (dX L(j,d)) denote the translation of j to the right and left, correspondingly, in d. Therefore, we can calculate the resulting translation of item j in space dimension d by:

$d X^{j, d}=d X_R^{j, d}+d X_L^{j, d}$     (11)

where, the value may be either positive or negative. Therefore, the following phrase can be used to quickly get the current coordinates of Object j in Dimension d:

$X^{j, d}=X_0^{j, d}+r_1 \times d X^{j, d}$     (12)

For the sake of keeping the algorithm's random search behaviour intact, where X 0(j,d) is the preceding position of object j in dimension d, and r 1 is a uniformly distributed chance sum in the range [0,1].

3.3.2 Repetition and updating parameters

In the first phase of the search process, the positions of all objects are initially generated at random. Each object's updated position is calculated at each iteration using Eqns. (4)-(11) and its fitness value. Each iteration requires a change to a system parameter, the stiffness coefficients of the springs, according to Eq. (4). The optimization problem determines the stopping criteria, which may be the maximum sum of iterations. Here is a quick rundown of how the spring search algorithm works:

3.3.3 Binary spring search procedure

Particle motion in a discrete search space can be applied with a binary formulation. Think of a hype-cube, the corners of which have coordinates that are a sequence of ones and zeroes. It is pointed out that there is a solution to the problem in each of the hyper-vertices. cube's In order to conduct a search in such a setting, the things being sought must "hype-cube hop" from one corner to another. Every particle's coordinate in discrete search space is either zero or one. Each particle is able to travel by switching the value of one or more dimensions between zero and one. To do this, we use a probability function to take into account the sum of displacement for each object in each dimension, and assume that the object moves somewhere along that dimension. Therefore, in the binary version of SSA, the term "dX(j,d)" refers not to the translation of object j in dimension d but rather to the probability that X(j,d) is either zero or one.

The updating mechanism, the formulas used to calculate the force applied to each object, and the amount of displacement each object experiences are all the same in the binary version of SSA as they are in the continuous version. For binary SSA, the probability function defined as dX(j,d) should be limited to the interval [0,1]. Because of this, the following expression is commonly used:

$S\left(d X^{j, d}(t)\right)=\left|\tanh \left(d X^{j, d}(t)\right)\right|$     (13)

The three-dimensional coordinates of an item are found by using the probability function as follows:

If $\operatorname{rand}<S\left(d X^{j, d}(t)\right)$ then,

$X^{j, d}(t+1)=$ complement $\left(X^{j, d}(t)\right)$     (14)

Else $X^{j, d}(t+1)=X^{j, d}(t)$.

The likelihood that item j will move in dimension d is given by the expression (14), and it increases as the value of dX(j,d) increases. In the interval [0, 1], rand is a uniformly distributed random number.

3.4 Proposed XGBoost algorithm

The data set is split. The training data is applied to the proposed XGBoost algorithm with BSSA model to classify the data. The test data will be tested using an already trained XGBoost model. Finally, it predicts jaundice disease accurately. Extreme Gradient Boosting (XGBoost) is an ensemble tree that follows a gradient descent structure to enhance the weak learners. It is a level by level additive modelling. Initially, XGBoost fits the data to the weak classifier. Later, it fits the data to another weak classifier for increasing the ith accuracy. The overall workflow is summarized as follows:

Let consider the input features DS  and it is represented in Eq. (15).

$D S=b: c, \quad|D S|=n, b \in R^m, c \in R$     (15)

where, n is the sum of occurrences, m denotes the sum of features, and b and c denote the features and target variable, correspondingly. The dataset contains n = 125 instances and m = 5 features. The sum of the predicted scores in k trees are calculated using Eq. (16).

$\widehat{c_i}=\sum_{k=1}^K f_k\left(b_i\right), \quad f_k \in F$      (16)

where, $\widehat{c_i}$  indicates the instance in the kth boost, $b_i$ indicates ith instance of training set.

The value of kth tree is represented by $f_k\left(b_i\right)$ and all the values of decision tree is represented by F. The XGBoost reduces the lose function $L_{\mathrm{k}}$ and it is given Eq. (17).

$L_k=\sum_{i=1}^n L\left(\widehat{c_i}, c_i\right)$     (17)

XGBoost defines the learning rate and hyper parameter values and it is given in Table 1.

Table 1. Parameters of proposed classifier

In XGBoost, the objective function considers the loss function and regularization term to choose predicate function. The Objective Function (OF) of XGBoost is given in Eq. (18).

$O F=\sum_{i=1}^n\left(\widehat{c}_i, c_i\right)+\sum_{i=1}^k R\left(f_i\right)$     (18)

where, L indicates the loss function, $\widehat{C_i}$ indicates predicted label and $c_i$ indicates actual label. R(f) indicates penalizing computing function in the training tree.

To define the complexity, we first define the tree function $f(b)$, which is given in Eq. (19).

$f(b)=w_q(b), w \in R^L, q: R^m \rightarrow\{1,2, \ldots L\}$     (19)

where, w indicates the leaf score of the tree, q indicates mapping function between data samples and corresponding leaf. The L indicates number of leaves in the tree.

The complexity of penalize model’s calculation is given in Eq. (20).

$R(f)=\gamma \times L+a(\|w\|)+\frac{1}{2} \times \lambda \times\|w\|^2$     (20)

where, λ and γ are constant values, each leaf value is represented by γ. $\|w\|$ indicates weight value of the tree.

The XGBoost follows an additive model which indicates the curve tree result with previous tree result. At t-th step, the Objective Function (OF) is calculated and it is given in Eq. (21).

$O F^{(T)}=\sum_{i=1}^n L\left(c_i, \hat{c}_i^t\right)+f_t\left(b_i\right)+R\left(f_t\right)+c$     (21)

where, $f_t$ indicates minimization of objective function and c is a constant.

Further, we compute the second order Taylor series and it is given in Eq. (22).

$O F^{(T)}=\sum_{i=1}^n\left[L\left(c_i, \hat{c}_i^{t-1}\right)+g_i \times f_t\left(b_i\right)\right.$$\left.+\frac{1}{2} \times h_i f_t^2\left(b_i\right)\right]+R\left(f_t\right)+c$     (22)

The calculation of $g_{\mathrm{i}}$ and $h_{\mathrm{i}}$ are given in Eqns. (23)-(24).

$g_i=\partial_{\hat{c}_i^{t-1}} L\left(c_i, \hat{c}_i^{t-1}\right)$    (23)

$h_i=\partial_{\hat{c}_i^{t-1}}{ }^2 L\left(c_i, \hat{c}_i^{t-1}\right)$    (24)

In order to remove constant, we obtain Eq. (12) by adding regularization term from Eq. (20).

$O F^{(T)}=\sum_{i=1}^n\left[g_i \times f_t\left(b_i\right)+\frac{1}{2} \times h_i f_t^2\left(b_i\right)\right]+\gamma^t$$+\alpha \times \sum_{j=1}^T w_i^2$     (25)

Finally, the XGBoost algorithm is selected good features and predicts the disease accurately. The overall XGBoost classification process is given in Algorithm 2.