A novel technique based on mutual information weighted feature selection to predict chronic kidney disease

Abstract

Chronic Kidney Disease (CKD) is a crucial life-threatening condition due to impaired kidney functionality and renal disease. In recent studies, Kidney disorder is considered one of the essential and deadliest issues that threaten patients’ survival with the lack of earlier prediction and classification. The earlier prediction process and the proper diagnosis help delay or stop the chronic disease progression into its final stage, where renal transplantation or dialysis is a known way of saving the patient’s life. Global studies reveal that nearly 10% of the population is affected by Chronic Kidney Disease (CKD), and millions die because of non-affordable treatment. Early detection of CKD from the biological parameters would save people from this crisis. Machine Learning algorithms are playing a predominant role in disease diagnosis and prognosis. This work generates compound features from CKD indicators by two novel algorithms: Correlation-based Weighted Compound Feature (CWCF) and Feature Significance based Weighted Compound Feature (FSWCF). Any learning algorithm is as good as its features. Hence, the features generated by these algorithms are validated on different machine learning algorithms as a test for generality. The simulation is done in MATLAB 2020a environment where various metrics like prediction accuracy gives superior results compared to multiple other approaches. The accuracy of CWCF over different methods like LR is 97.23%, Gaussian NB is 99%, SVM is 99.18%, and RF is 99.89%, which is substantially higher than the approaches without proper methods feature analysis. The results suggest that generated compound features improve the predictive power of the algorithms.

Keywords

Feature selection correlation feature significance chronic kidney disease feature projection mutual information

1 Introduction

In endeavoring to work on healthcare by employing computational techniques, extracting relevant and appropriate features is essential. The reduction of high-dimensional data plays a crucial preprocessing method while handling healthcare data analytics [1]. Processing features without a reduction of dimensionality may increase the processing and computational time. Hence optimized feature extraction and selection from high dimensional data has a more significant impact on classification techniques [2]. It is hardly known ahead of time which features are appropriate for classification. Also, it is not practically possible to measure the features at the initial stage that are to be deployed during classification. Extracting meaningful and significant features will reduce the computational complexity, increase the model’s reliability to classify effectively, and ensure precise model representation [3]. Predominantly there are two types of high dimensionality reduction techniques that will optimize the features to a greater extent, namely feature extraction and feature selection [4, 5]. Feature extraction works on feature scaling and transformation that bring all the features on one scale by highlighting the high dimensional feature data to lower subspace which will help to maintain the quality of the feature to be the original and minimize the loss of information during reduction.

Feature selection points directly to the subset of features primarily important from the extracted features for classification. Hence feature extraction and feature selection for dimensionality reduction has become one of the bases for the classifiers. Considering feature extraction’s importance, there are different approaches, namely principal component analysis [6], Isometric mapping, and locally linear embedding. In contrast, feature selection is commonly constructed on information theory, statistics, and rough sets classified depending on the protocols [7]. In the early prediction of medical diseases, feature selection plays a significant role in identifying the essential features needed for diagnosis [8]. Hence, selecting appropriate features helps narrow down the medical expert’s decision during diagnosis. Also, the healthcare dataset contains large voluminous quantities of medical records that are directly taken from the hospitals and are updated daily. These datasets constitute information from expert prescriptions diagnostics reports and posts on social networking like tweets, tags and blogs [9, 10]. The origin of emerging techniques like data mining, machine learning, and intelligent advent technologies like the Internet of Things and Edge computing has given a new outset in diagnosing Chronic Kidney Disease with great force to a new epoch.

Medical analysis has become a remunerative research area that focuses on diagnosing chronic diseases like kidney, heart and diabetes [11]. Developing a proper Chronic Disease Diagnosis system called CDD can be used as a common framework for continuous assessment of chronic disease in affected persons [12]. Hence, developing an automated system with appropriate feature selection aid helps medical physicians give accurate early predictions of disease. Feature selection is also investigated as variable selection. It is a regularly performed preprocessing method that eliminates redundant attributes and irrelevant and missing values and processes on insignificant and surplus data [13]. These processes would enhance the computational speed, accelerate clear visualization of results, propagate exploratory analysis of data, and ensure high accuracy and performance of prognosis of the disease [14, 15]. Various popular selection methods, such as ensemble methods, are more popularly known for combining multiple hypotheses to get a better hypothesis. Wrapper methods are used explicitly on machine learning algorithms and embedded methods more like wrapper methods that give optimized output and filter methods that work on independent features [16]. These methods are more frequently used in prevalent chronic diseases such as kidney, heart, diabetes, stroke and thalassemia [17]. It is proved by developing a highly meticulous diagnostic system to predict joint malfunctions by deploying VAG signals that describe the importance of feature selection in the medical field [18]. Thus, there is a need for an efficient approach to deal with the feature learning and classification process. The significant research contributions are provided below:

Initially, acquire the samples from available online data sources like publicly accessible datasets. Here, the UCI ML repository is considered for CKD analysis.

To provide a detailed description of the linear feature projection with correlation-based weighted compound features, the mutual information among the features is analyzed.

The experimental outcomes are discussed where the simulation is done with MATLAB 2020a environment, and various metrics like precision, recall, F1-score and accuracy are measured and compared. Based on this analysis, it is proven that the model outperforms multiple existing approaches.

This research paper is organized as follows: Section 2 discusses the various existing approaches with the significant research constraints associated with the prediction process. It paves the baseline for the proposed methodology. Section 3 gives a short description of linear feature projection for finding mutual information between the target class and feature for chronic kidney disease. Secondly, Correlation Based Weighted Compound Feature Selection Using Linear Projection to perform mathematical estimation is presented. Further, an algorithm by combination with linear projection is given. Thirdly, Feature Significance Based Weighted Compound Feature Selection Using Linear Projection with the deployment of a partial feature set for chronic disease prediction is shown in Section 4. Section 5 gives a discussion about the results for the prediction of chronic kidney diseases is presented. Lastly, a comparison of different ML algorithms is discussed with the evaluated results as shown in the conclusion.

2 Related works

Various existing approaches are discussed to predict CKD in its earlier stage. Tangri et al. [23] depict NCA for kidney cancer classification of sub-types with miRNA genome data. The provided miRNA sample classification over cancer sub-types is offered to acquire discriminative properties of LSTM and miRNAs in the RNN form. This anticipated NCA process chooses the biased miRNAs dataset. The miRNA subset facilitates the anticipated LSTM to cluster the cancer-based miRNA to five diverse sub-types with 95% average prediction accuracy and 0.93 MCC for ten random clustered processes (five times) that are nearer to the average output of miRNA for rating the functionality. Sinha et al. [24] initiated a novel DNN approach to classify and detect various renal histologic phenotypes. It is revealed that the ML concept with DNN has provided extensive performance in various processing tasks of histologic images. The NN concepts intend to extract and use quantitative feature learning to categorize diversions among different genotypes (mice). The non-glomerular segmentation and animal-based genotype rely on the quantitative image features and pretend to provide superior performance.

Jena et al. [25] recommend HNN for predicting kidney function. A prediction problem is explicitly formulated as a binary classification function to acquire the complete information from the available Electronic Health Record (EHR). The author recommends HNN for integrating bi-directional LSTM and auto-encoder network model. Based on the data acquired from the raw HER, the author initiates the construction of a dataset and collects roughly 35k reports from various patients’ including patients with hypertension. The testing outcome shows that 90% of the model accuracy is due to the inclusion of NN. The anticipated approach has examined the need for synthetic NN with the predictions for various analyses. The efficacy of the anticipated NN model relies on the construction of an available online dataset. Author et al. [26] model a novel RNN for predicting acute kidney injury development. The model determines the input sequences responding to various prior sequential sections essential for further processing. The prediction probabilities of acute kidney disease with specific clinical parameters/sequences are predicted with RNN. For all acute kidney disorders defined with crucial severity levels, the provided model intends to show reasonable prediction with 55% for all acute kidney diseases. The advancements with the clinical diagnosis in an average time interval of about 48 hours at the selected operation give 35% accuracy. Acute kidney disease is identified for one patient among three different predicted cases, while some other predictions are estimated to be wrong. Moreover, testing reveals that the patients with CKD show 58% FP evaluation.

Kolachalama et al. [27] anticipated an ensemble-based multi-stage DL approach for segmenting kidney tumors. An integration process is applied to analyze the prediction outcomes from the previous phases and to measure the variance among diverse individual models. The dice score (average) for kidney disease and tumor cases are 96% and 74%, respectively, for some unidentified test cases. Some predictions are positive and enhanced by providing advanced knowledge of benign cysts, which decreases tumor segmentation frequently over the IoMT platform. The numerical result drastically diminishes the error, and a relatively large batch size is used. The batch normalization properties are exploited superiorly, contrary to various attempts with smaller batch sizes. The outcomes acquired from 32 samples are enhanced among the evaluations. However, all these models fail to give superior prediction accuracy based on earlier feature analysis. Thus, the drawback needs to be addressed, and this research has initiated an attempt to perform better feature learning for the earlier prediction process. Table 1 depicts the comparison of various existing approaches and their corresponding drawbacks.

Table 1
Comparison of various existing approaches

References Methods Approach Results

Tangri et al., [23] Automated sensing of CKD Sampling and sensing of incoming data using learning approaches Better prediction is made with a glomerular filtration rate

Sinha et al., [24] CKD prediction k-NN, RF, and NB classifier model The new selection process is done with CKD prediction, where RF and NB are executed for CKD prediction.

Lazich et al., [26] CKD prediction with SVM for feature selection SVM Two diverse evaluators are used for filtering, and the greedy step-wise searching engine gives a nominal prediction rate.

Sheehan et al., [28] CKD prediction using predictive attributes NN and wrapper-based embedding model CKD prediction with k-NN, RF, and NB classifiers gives an average prediction rate that ranges from 80% to 90%

Kallenberger et al., [29] CKD analysis SVM with simulated program Sensitivity analysis with SVM classifier which gives a prediction accuracy of 93% with linear kernel

References	Methods	Approach	Results
Tangri et al., [23]	Automated sensing of CKD	Sampling and sensing of incoming data using learning approaches	Better prediction is made with a glomerular filtration rate
Sinha et al., [24]	CKD prediction	k-NN, RF, and NB classifier model	The new selection process is done with CKD prediction, where RF and NB are executed for CKD prediction.
Lazich et al., [26]	CKD prediction with SVM for feature selection	SVM	Two diverse evaluators are used for filtering, and the greedy step-wise searching engine gives a nominal prediction rate.
Sheehan et al., [28]	CKD prediction using predictive attributes	NN and wrapper-based embedding model	CKD prediction with k-NN, RF, and NB classifiers gives an average prediction rate that ranges from 80% to 90%
Kallenberger et al., [29]	CKD analysis	SVM with simulated program	Sensitivity analysis with SVM classifier which gives a prediction accuracy of 93% with linear kernel

3 Methodology

This section provides a detailed analysis of the anticipated feature learning approach where the evaluation is done with MATLAB 2020a simulation environment. Various performance metrics like accuracy, precision, F-measure, and recall are evaluated and compared with existing approaches. Figure 1 depicts the block diagram of the anticipated model.

Fig. 1

Block of the proposed predictor model.

3.1 Dataset acquisition

This research considers the online available UCI ML dataset for the CKD dataset, and the UCI dataset contains 24 attributes collected from 400 patients. Table 2 summarizes the attributes in the dataset.

Table 2
Attributes in the CKD dataset

Feature Range Feature Range

Age 2-90 years Serum creatinine 0.4-76

Blood pressure 50-180 Sodium 4.5-163

Specific Gravity 1.005- 1.025 Potassium 2.5-47

Albumin 0-5 Hemoglobin 10.3-17.8

Sugar 0-5 Packed cell volume 9-54

Red Blood Cells Yes, no White blood cell (count) 2200-26400

Pus cells Yes, no Red blood cell (count) 2.1-8

Pus cell clumps Yes, no Hypertension Yes, no

Bacteria Yes, no Diabetes Mellitus Yes, no

Blood glucose random 22-490 Coronary Heart disease Yes, no

Appetite Yes, no Blood urea 1.5-391

Anemia Yes, no Pedal enema Yes, no

Feature	Range	Feature	Range
Age	2-90 years	Serum creatinine	0.4-76
Blood pressure	50-180	Sodium	4.5-163
Specific Gravity	1.005- 1.025	Potassium	2.5-47
Albumin	0-5	Hemoglobin	10.3-17.8
Sugar	0-5	Packed cell volume	9-54
Red Blood Cells	Yes, no	White blood cell (count)	2200-26400
Pus cells	Yes, no	Red blood cell (count)	2.1-8
Pus cell clumps	Yes, no	Hypertension	Yes, no
Bacteria	Yes, no	Diabetes Mellitus	Yes, no
Blood glucose random	22-490	Coronary Heart disease	Yes, no
Appetite	Yes, no	Blood urea	1.5-391
Anemia	Yes, no	Pedal enema	Yes, no

3.2 Linear feature projection

The classifiers are spawning over a multitude of features to derive reliable hypotheses. Many dimensionality reduction techniques like Principal Component Analysis (PCA), Wrappers, and filters are commonly used to:

Eliminate redundancy

Explore high dimensional more information from features

Generalizing the classifier

The maximum discriminative information is preserved between features to obtain a more generic classification. Independent Component Analysis (ICA) is essential to find more appropriate features. Let y = Ax be the maximum informative linear feature projection from m to n dimensions. The projection matric A ∈ R^m*n is optimally approximated by accumulating the mutual information M_info (f|t) between features f and target class t [19]. It is mathematically expressed as in Equation (1):

$M_{info} (f | t) \approx \sum_{i = 1}^{n} M_{info} {(f_{i} | t)}_{i}$ (1)

From Equation (1), the aggregation of mutual information between f and t is provided. The projection of features associated with the target class can be ranked by the individual features’ M_info (f|t). The entropy measure is maintained to preserve the orthogonality. Conditional entropy is estimated when the features are independent of one another. It is the uncertainty introduced by a feature when correlated with another feature. This uncertainty can be interpreted as the inter-information provided by a feature on another feature [20]. Internally, the estimation of M_info (f|t) is grounded on the entropy of the independent features with its target class. The individual entropy of the feature H (f_i) based on the m space estimator is given by Equation (2):

$\begin{matrix} H (f_{i}) \\ = - \frac{1}{N - m} \sum_{i = 1}^{N - m} log (N + 1) (f_{i + m} - f_{i}) / m \end{matrix}$ (2) The entropy of the features concerning the target class t is found according to Equation (3):

$\begin{matrix} H (f_{i} | t) = & - \frac{1}{N_{t} - m_{t}} \sum_{i = 1}^{N_{t} - m_{t}} log (N_{t} + 1) \\ (f_{i + m_{t}} - f_{i}) / m_{t} \end{matrix}$ (3)

Finally, the estimation of M_info (f|t) happens according to Equation (4), where P_t is a constant.

$M_{info} (f_{i} | t) = H (f_{i}) - \sum_{t} P_{t} H (f_{i} | t)$ (4)

3.3 Correlation based weighted compound feature selection using linear projection

Correlation is the mathematical estimation of the intensity of association between the feature and the target class. High correlation indicates that the features play a predominant role in prediction. Weighing the features based on their correlation enhances the efficacy of prediction. Generating compound features from the original features must preserve the orthogonality among new combinations of features and the original features. The proposed correlation-based weighted compound feature selection uses linear projection methodology and then imposes correlation as weights. The Partial Features (PF) consist of features that have M_info (f|t) value below the threshold limit (T₁). The Subset of Semi Features (SSF) set is constructed using these weighted features and is given in Equation (5):

$SSF \leftarrow c_{1} f_{1} + c_{2} f_{2} + . . . + c_{n} f_{n}$ (5)

The correlation values are represented as C ={ c₁, c₂, …, c_n }. The ultimate Correlation-based Weighted Compound Features (CWCF) set is constructed with features that exceed the qualifying thresholds, computed empirically. Algorithm 1 describes the entire process of Correlation-based feature selection grounded on a linear projection of features.

Algorithm 1: Correlation based weighted compound feature selection using linear projection
Input: Correlation C ={ c₁, …, c_n }, master feature set MF ={ f₁, f₂, …, f_n } ;
Output: Correlation based Weighted Compound Features (CWCF)
1. Map mc: f₁ → c_i where f_i ∈ MF and c_i ∈ C
2. Initialize PF→ ∅ //Partial Features (PF)
3. D_max (MF, PF)
4. Initialize CWCF [d]← ∅ ;
5. For f_i ∈ MF where i ∈ [1, N]
6. Compute $H (f_{i}) = - \frac{1}{N - m} \sum_{i = 1}^{N - m} log (N + 1) (f_{i + m} - f_{i}) / m$ as in Equation (2);
7. Compute $H (f_{i} \| t) = - \frac{1}{N_{t} - m_{t}} \sum_{i = 1}^{N_{t} - m_{t}} log (N_{t} + 1) (f_{i + m_{t}} - f_{i}) / m_{t}$ as in Equation (3);
8. Evaluate M_info (f_i\|t) = H (f_i) - ∑_tP_t H (f_i\|t) as in Equation (4);
9. End for
10. for each M_info (f_a\|t) < T₁ in MF;
11. PF ← PF ∪ { f_a } ;
12. MF ← MF - { f_a } ;
13. End for
14. For each (PF_a > 0) in PF
15. SSF = c_a* f_a ;
16. End for
17. CWCF - MF ∪ SSF ;
18. Return CWCF

4 Feature significance based weighted compound feature selection using linear projection

Feature importance is a score assigned to features based on their usefulness in predicting the target class. The information gain-based feature selection is widely used as it dramatically reduces dimensions without loss of information [21]. The proposed algorithm used model agnostic feature significance to construct compound features from the partial feature set. The construction of a partial feature set is governed by Equation (6):

$SSF \leftarrow s_{1} f_{1} + s_{2} f_{2} + . . . + s_{n} f_{n}$ (6)

From Equation (6), s₁, s₂, . . , s_n represents the feature significance score concerning the target class t. Algorithm 2 shows the weighted compound feature selection generation using entropy-based linear projection with the feature significance scores as weights.

Algorithm 2: Feature Significance based weighted compound feature selection using linear projection
Input: Feature Significance S ={ s₁, s₂, . . , s_n } and master feature set MF ={ f₁, f₂, …, f_n }
Output: Feature Significance based Weighted Compound Features (FSWCF)
1. sMap mc : f_i → s_i where f_i ∈ MF and s_i ∈ S
2. Initialize PF→ ∅ //Partial features (PF)
3. d ← max (MF, PF)
4. Initialize CWCF [d]→ Φ ;
5. For f_i ∈ MF where i ∈ [1, N]
6. Evaluate $H (f_{i}) = - \frac{1}{N - m} \sum_{i = 1}^{N - m} log (N + 1) (f_{i + m} - f_{i}) / m;$
7. Evaluate $H (f_{i} \| t) = - \frac{1}{N_{t} - m_{t}} \sum_{i = 1}^{N_{t} - m_{t}} \frac{log (N_{t} + 1) (f_{i + m_{t}} - f_{i})}{m_{t}};$
8. Evaluate M_info (f_i\|t) = H (f_i) -∑_tP_t H (f_i\|t) ;
9. End for
10. for each M_info (f_a\|t) < T₁ in MF;
11. PF ← PF ∪ { f_a } ;
12. MF ← MF - { f_a } ;
13. End for
14. For each (PF_a > 0) in PF
15. SSF = s_a* f_a ;
16. End for
17. FSWCF - MF ∪ SSF ;
18. Return FSWCF

The selected features need to be optimized using any meta-heuristic optimizer to verify the global and local outcomes. The model needs to fulfill both exploration and exploitation. Here, a novel Rule-based firefly optimizer is proposed to analyze the significance of selected features.

In accordance with the no-free lunch theorem, the averaged objective function of any ML algorithm is as good as the others. So, the proposed work delves into selecting the features which hold higher prominence, than exploring all the features. The proposed work significantly reduces the overhead of the learning algorithm by presenting a subset of significant features with appropriate weights deduced based on the correlation with the target. This scheme offers a two-fold advantage: firstly, it reduces the processing overhead, and secondly, it confines the algorithm to learn only from the significant features. This is in contrast to the existing methods in the literature which are limited to selecting the features from Exploratory Data Analysis. Also, forming subsets of features based on mutual information, significance and correlation in the healthcare domain gives detailed knowledge about the combination of factors that worsens the health condition.

5) Rule-based Firefly optimizer

The target of this optimizer is to attain an optimal solution and enhances the prediction rate [31]. It is a heuristic optimizer energized with the flashing nature of fireflies and works based on the signal system. The model follows certain rules: Rule 1- enables an individual to attract other fireflies; Rule 2- attractiveness is directly proportional to brightness, i.e. less brightness is attracted towards brighter one; Rule 3- firefly moves randomly and determined with objective functions. The light intensity variation is expressed as in Equation (7):

$I (d) = I_{0} e^{- γ d^{2}}$ (7)

Here, γ specifies light absorption rate; I_o specifies initial light intensity. The firefly attractiveness (rule 2) is expressed as in Equation (8):

$β (d) = β_{0} e^{- γ d^{2}}$ (8)

Here, d specifies distance among fireflies; β₀ specifies distance attractiveness and γ specifies variation attractiveness. Finally, rule 3 is measured with Cartesian distance measure as in Equation (9):

$d_{ij} = \sqrt{\sum_{k = 1}^{n} {(x_{ik} - x_{jk})}^{2}}$ (9)

Here, i and j specify the distance between two fireflies, x_ik specifies the spatial coordinator of k^th component; N specifies the number of dimensions. The movement (random) of fireflies is expressed as in Equation (10):

$\begin{matrix} x_{j} (v + 1) = & x_{i} + β_{o} e^{- γ d^{2}} (x_{j} - x_{i}) \\ + α (rand - \frac{1}{2}) \end{matrix}$ (10)

Here, α specifies randomization parameter and d specifies the random number distributed uniformly. Algorithm 1 explains the Rule-based firefly functionality for optimization:

Algorithm 3 Rule-based firefly optimizer
Input: Maximum generation, population size;
Output: Best solution
1. Initialization: firefly population x (i = 1, 2, …, n)
2. Evaluate light absorption value
3. do
4. for i = 1, 2, 3, . . . n ;
5. for j = 1, 2, . . . i ;
6. predict the light intensity by f (x_i)
7. if light intensity is higher than others then
8. Move firefly i towards j direction based on corresponding dimensions
9. Endif
10. Measure attractiveness variation with d distance;
11. Compute $Δ f = f (x_{i}^{*}) - f (x_{i})$
12. Compute $U = exp (- \frac{Δ f}{T_{m}})$
13. if Δf ⩽ 0 or U > R (0, 1)
14. $X_{i} \leftarrow X_{j}^{*}$
15. End if
16. Compute random using control parameter;
17. compute new solution and re-new light intensity;
18. end for
19. end for
20. Classify every fly and achieve best solution;
21. Perform stopping criteria; // when condition achieved.

5 Numerical results and discussion

This section provides the numerical analysis of the anticipated model over various existing approaches. Here, MATLAB 2020a is used as a simulation environment where metrics like accuracy, precision, recall, and F1-measure are evaluated and compared with existing approaches. The metrics are mathematically expressed as in Equations (10): $Accuracy = \frac{TP + TN}{TP + TN + FP + FN}$ (11) $Precision = \frac{TP}{TP + FP}$ (12) $Recall = \frac{TP}{TP + FN}$ (13) $F - measure = \frac{2 (TP)}{2 ((TP) + (FN))}$ (14)

Here, TP is True positive; TN- True Negative; FP- False Positive; FN –False Negative [32]. Based on the analysis, it is evident that medical data is highly unstructured and selecting representative features for developing an accurate prediction model is a tiresome task. Since medical data is susceptible, additional care has been taken to preprocess the data and select pertinent features. The primary objective of this work is to assess the factors that contribute to CKD among the features collected in the benchmarked Chronic Kidney Disease dataset from the UCI machine learning repository [22]. The proposed algorithms, namely Correlation Based Weighted Compound Feature Selection Using Linear Projection and Feature Significance based weighted compound feature selection using Linear Projection, aim to generate the best compound features from the original master feature set. The so-developed compound features are applied over the prediction algorithms, namely SVM, random forest, logistic regression, and Gaussian Naïve Bayes, to test its effectiveness in predicting CKD occurrence among patients. The correlation-based algorithm produces compound features to classify whether a patient is prone to CKD. The correlation values of each feature with the target class are listed in Table 2. The dataset attributes (features) considered for analysis include age, BP, gravity, albumin, sugar, RBC, pus cells, clumps, bacteria, glucose (random), appetite, creatinine, sodium, potassium, hemoglobin, hypertension, diabetes mellitus, coronary heart disease, urea, enema, anemia, RBC, WBC and packed cell volume are included. However, all these features are not highly required for complete validation. Some influencing features are considered for analysis where some analysis metrics like correlation, feature significance and MI are analyzed, and the corresponding values are evaluated. The feature significance value relies on 0 to 2, the MI relies on 0 to 0.45, and the correlation value includes positive and negative values.

The applications in the healthcare domain hold accuracy as its primary metric rather than time consumption. The proposed method investigates all the features based on their mutual information, feature significance and correlation to ascertain its rationale in predicting CKD. Also, the numeric nature of clinical parameters further speeds up the computation as present-day systems are well equipped with enormous processing power.

The compound features consist of original features and SSF to predict CKD occurrence effectively. Table 3 summarizes the selection of features by both the proposed algorithms. Based on Table 3, the master features include f₁ → f₆, partial features with lesser mutual information have f₂ → f₄, the semi-subset features are f₁, f₅ and f₆, and the semi subset features FSWCF includes f₁, f₅and f₆ .

Table 3

Assessment of various features in the context of Correlation, Feature Significance and Mutual Information

Feature	Correlation	Feature Significance	Mutual Information
Age	0.225405	0.00200	0.04040996
Blood pressure	0.290600	0.00364	0.19098851
Specific Gravity	-0.698957	0.13081	0.34808052
Albumin	0.599238	0.07111	0.11736615
Sugar	0.327812	0.00525	0.13867268
Red Blood Cells	0.474070	0.02276	0.11258548
Pus cells	0.420356	0.00447	0
Pus cell clumps	0.265313	0.00021	0.00659575
Bacteria	0.186871	0	0.16989665
Blood glucose random	-0.755806	0.04371	0.20864111,
Appetite	-0.393341	0.00437	0.07527314
Serum creatinine	0.294079	0.07664	0.26010843
Sodium	-0.342288	0.00606	0.22065426
Potassium	0.076921	0.00296	0.47251081
Hemoglobin	-0.729628	0.17311	0.43571206
Hypertension	0.590438	0.02672	0.17823579
Diabetes Mellitus	0.559060	0.02562	0.01305714
Coronary Heart disease	0.236088	0	0.12345069
Blood urea	0.372033	0.00854	0.38319485
Pedal enema	0.375154	0.00028	0.12559627
Anemia	0.325396	0.16678	0.45077862
Packed cell volume	0.172033	0.14703	0.10793548
RBC	0.215154	0.07362	0.40437896
WBC	0.072033	0.00281	0.22246756

5.1 Discussion of CWCF

The proposed algorithm 1 generates CWCF, used by various machine learning algorithms such as SVM, Gaussian Naïve Bayes, Logistic Regression and Random Forest to test its efficacy. The mean of M_info (f|t) is considered a threshold for simplicity. Various techniques like Ant Colony Optimization (ACO) and Particle Swarm Optimization (PSO) can be explored to determine the threshold value. Detailed correlation values and mutual information is listed in Table 2. The train-test ratio of 75-25 is preserved throughout the experiment. The performance of various algorithms on the generated CWCF is assessed on the metrics like classification accuracy, F1-score, Precision and Recall. Table 4 gives the performance results of the CWCF features selection algorithm on standard machine learning algorithms. The impact of compound feature generation is well portrayed in the results.

Table 4
Summary of feature selection by proposed algorithms

Master features f₁, f₂, f₃, f₄, f₅, f₆

Partial features with low M_info (f|t) f₂, f₃, f₄

A subset of semi-features (algorithm1) SSF ← c₂ * f₂ + c₃ * f₃ + c₄ * f₄

CWCF { SSF , f ₁, f ₅, f ₆}

A subset of semi-features (algorithm2) SSF ← s₂ * f₂ + s₃ * f₃ + s₄ * f₄

FSWCF { SSF , f ₁, f ₅, f ₆}

Master features	f₁, f₂, f₃, f₄, f₅, f₆
Partial features with low M_info (f\|t)	f₂, f₃, f₄
A subset of semi-features (algorithm1)	SSF ← c₂ * f₂ + c₃ * f₃ + c₄ * f₄
CWCF	{ SSF , f ₁, f ₅, f ₆}
A subset of semi-features (algorithm2)	SSF ← s₂ * f₂ + s₃ * f₃ + s₄ * f₄
FSWCF	{ SSF , f ₁, f ₅, f ₆}

Figure 2 visualizes the results of Table 4. The machine learning algorithms exhibit better performance when subjected to features generated by CWCF. Logistic regression shows improved accuracy of 97.23% when deployed with features selected through the CWCF method against raw features. A similar trend can be observed with other algorithms also. In addition to this, the CWCF reduces misclassifications, which are very crucial in any medical disease diagnosis. It is witnessed by the improvement in the F1- Score, precision and recall. Thus by selecting influential features through CWCF, the ML algorithms can output substantially improved efficacy, which is the need of the hour for highly unstructured heterogeneous medical data. In the case of original feature analysis, the prediction accuracy of RF is 98.65% which is 2.82%, 0.54% and 0.005% higher than LR, NB and SVM. The F1-score of RF, NB, and SVM is 98% which is 2% higher than LR. The precision of NB, SVM, and RF is 98% which is 3% higher than NB, SVM, and RF approaches. The recall of RF, NB and SVM is 98% which is 1% higher than LR. From this, it is known that the models like RF, NB, SVM and LR give nominal outcomes for original features. While in the case of CWCF, the prediction accuracy of RF is 99.89% which is 2.66%, 0.76%, and 0.71% higher than LR, NB, and SVM. The F1-score of RF is 99.13% which is 0.13% higher than NB and SVM and 1.13% higher than LR. The precision of RF is 99.13% which is 0.13% higher than NB and SVM and 1.13% higher than LR. The value is the same for recall also. The CWCF model gives comparatively higher values than the original features.

Fig. 2

Competitive analysis of CWCF selection on various algorithms.

Table 6

Results of the proposed FSWCF feature selection algorithm applied to various machine learning techniques

Algorithm	Original Features				FSWCF
Accuracy	F1-Score	Precision	Recall	Accuracy	F1-Score	Precision	Recall
Logistic Regression	92	92	97	89	93.94	94	95	94
Gaussian Naïve Bayes	97.32	94.23	95	91	99.24	99	100	98
SVM	95	95	99	93	96.21	96	97	98
Random Forest	98.32	94	93	95	99.89	99	99	99

With the mutual comparison, the prediction of RF is 99.89%, SVM is 99.18%, NB is 99.13%, and LR is 97.23% which is 1.24%, 0.58%, 1.07% and 3.95% higher than RF, SVM, NB and LR-based original feature analysis. The F1-score of RF is 99.13%, SVM is 99%, NB is 99%, and LR is 98% which is 1.13%, 1%, 1% and 3% higher than RF, SVM, NB and LR-based original feature analysis. The precision of RF is 99.13%, SVM is 99%, NB is 99%, and LR is 98% which is 1.13%, 1%, 1% and 3% higher than RF, SVM, NB and LR-based original feature analysis. The recall of RF is 99.13%, SVM is 99%, NB is 99%, and LR is 98% which is 1.13%, 1%, 1% and 1% higher than RF, SVM, NB and LR-based original feature analysis. Based on this analysis, it is proven that CWCF works better than the actual feature analysis.

5.2 Discussion on FSWCF

Algorithm 2 deploys feature significance as a primary criterion in selecting features from the master feature set. FSWCF algorithm considers significance scores as weighing metrics for weaker features with low M_info (f|t) values. These scores are generated by training the original features with the Random Forest algorithm, one of the best bagging algorithms aiming to reduce overfitting. The performance analysis of the CKD data in predicting the disease on deploying FSWCF compound feature generation on various machine algorithms is given in Table 5. Figure 3 shows the visual performance comparison of the ML algorithms.

Table 5
Results of the proposed CSCF feature selection algorithm applied to various machine learning techniques

Algorithm Original Features CWCF

Accuracy F1-Score Precision Recall Accuracy F1-Score Precision Recall

Logistic Regression 95.83 96 95 97 97.23 98 98 98

Gaussian Naïve Bayes 98.11 98 98 98 99.13 99 99 99

SVM 98.6 98 98 98 99.18 99 99 99

Random Forest 98.65 98 98 98 99.89 99.13 99.13 99.13

Algorithm	Original Features	CWCF
Logistic Regression	95.83	96	95	97	97.23	98	98	98
Gaussian Naïve Bayes	98.11	98	98	98	99.13	99	99	99
SVM	98.6	98	98	98	99.18	99	99	99
Random Forest	98.65	98	98	98	99.89	99.13	99.13	99.13

Fig. 3

Competitive analysis of FSWCF selection on various algorithms.

A significant improvement is seen when the ML algorithms are subjected to features selected through the FSWCF method than processing with original features. A steep improvement is sensed in Gaussian Naïve Bayes, where the accuracy improved by 2%, and the misclassification is reduced further. This technique proves its strength in reducing misclassifications rather than just tuning only the classification accuracy. From Table 5, the original feature analysis is done with metrics like accuracy, F1-score, precision and recall. The accuracy of RF is 98.32% which is 6.32%, 1% and 3.32% higher than LR, NB and SVM. The F1-score of SVM is 95% which is 3%, 0.77% and 1% higher than LR, NB and RF. The precision of SVM is 99% which is 2%, 4% and 6% higher than LR, NB and RF approaches. The recall of RF is 95%, which is 6%, 4%, and 2% higher than LR, NB, and SVM techniques. This analysis identifies that SVM works well compared to LR, NB and RF approaches. While in the case of FSWCF, the prediction accuracy of RF is 99.89% which is 5.95%, 0.65%, and 3.68% higher than LR, NB and RF. The F1-score of RF and NB is 99% which is 5% and 3% higher than LR and SVM. The precision of NB is 100% which is 5%, 3% and 1% higher than LR, SVM and RF approach. The recall of RF is 99% which is 1% higher than NB and SVM and 5% higher than LR.

With the mutual comparison, the prediction of RF is 99.89%, SVM is 96.21%, NB is 99.24%, and LR is 93.94% which is 1.57%, 1.21%, 1.92% and 1.94% higher than RF, SVM, NB and LR-based original feature analysis. The F1-score of RF is 99%, SVM is 96%, NB is 99%, and LR is 94% which is 5%, 1%, 1.68% and 2% higher than RF, SVM, NB and LR-based original feature analysis. The precision of RF is 99%, SVM is 97%, NB is 100%, and LR is 95% which is 6%, 2%, 5.77% and 3% higher than RF, SVM and NB LR-based original feature analysis. The recall of RF is 99%, SVM is 98%, NB is 98%, and LR is 94% which is 0.68%, 3%, 0.68% and 2% higher than RF, SVM, NB and LR-based original feature analysis. Thus, the functionality for feature selection with FSWCF is superior to the standard feature analysis.

Though both methods show a considerable increase in performance, the algorithm1 (CWCF) selection strategy proves to be more effective than the algorithm2, based on feature significance scores. The comparative assessment between their mean classification metrics is shown in Fig. 4. The improved performance of CWCF is because the correlation values are estimated based on the number of features in positive and negative classes. Any change in the numbers will alter the values. Hence this score is more reliable and specific to the dataset. On the other hand, the FSWCF considers the ranking of features based on their influence in labeling the classes. Also, the FSWCF needs the support of a good predictor to output enhanced performance. This work mainly focuses on building good compound features by deploying two techniques, namely CWCF and FSWCF. A rational improvement is sensed in using these methods. This work provides lots of scopes to train a dataset with a lesser number of features and to introduce evolutionary algorithms in optimizing the threshold values to select the features.

Fig. 4

Comparison between proposed methods.

While comparing the performance of FSWCF and CWCF, it is evident that FSWCF works well compared to CWCF approaches. The prediction accuracy of RF (FSWCF and CWCF) is 99.89%; however, the prediction accuracy of SVM is 96.21%, NB is 99.24%, and LR is 93.94% which is 2.97% lesser than SVM (CWCF) and 0.09% higher than NB (FSWCF) and 3.29% lesser than LR (CWCF). The F1-score (FSWCF) of RF is 99% which is 0.13% higher than RF (CWCF), SVM is 96% which is 3% lesser than SVM (CWCF), 99% for NB and 4% lesser than LR (CWCF). The precision (CWCF) is 99.13% which is 0.13% lesser than RF (FSCWF), 2% higher than SVM (FSCWF), 1% lesser than NB (CWCF) and 3% higher than LR (FSCWF), and recall of RF is 0.13% higher than RF (FSWCF), 1% higher than SVM (FSWCF), 1% higher than NB (FSWCF), and 4% higher than LR (FSWCF). It is proven that CWCF shows superior outcomes compared to other feature selection approaches with the optimal observed points using a rule-based firefly optimizer as in Fig. 5.

Fig. 5

Objective function of rule-based firefly.

6 Conclusion

Chronic Kidney Disease has become a prevalent global health issue. A variety of factors are found to create deterioration in the performance of kidney function, causing them to fail over a while. This research focuses on restricting the influence of various factors that contribute to CKD by assessing the attributes through two techniques, namely CWCF and FSWCF. The results of these techniques are evaluated by testing them with different ML algorithms, which shows that CWCF yields improved results over the FSWCF algorithm. The proposed model delivers better results compared to standard analysis. This research will help medical practitioners to decide whether the patient is prone to CKD. The primary research constraint is the analysis with the feature learning process alone without much concentration towards the classification process. In the future, the classification process will be done with ML approaches to enhance prediction accuracy.

References

Murthy

C.A.

, Bridging feature selection and extraction: compound feature generation, IEEE Transactions on Knowledge and Data Engineering 29(4) (2016), 757–70.

Raymer

M.L.

, Punch

W.F.

, Goodman

E.D.

, Kuhn

L.A.

and Jain

A.K.

, Dimensionality reduction using genetic algorithms, IEEE Transactions on Evolutionary Computation 4(2) (2000), 164–71.

, Wang

, Yang

, Liang

, Shang

and Wan

, Fast hybrid dimensionality reduction method for classification based on feature selection and grouped feature extraction, Expert Systems with Applications 150 (2020), 113277.

Choi

, Shin

J.H.

, Lee

, Sheridan

and Lu

W.D.

, Experimental demonstration of feature extraction and dimensionality reduction using memristor networks, Nano Letters 17(5) (2017), 3113–8.

Daza-Santacoloma

, Arias-Londono

J.D.

, Godino-Llorente

J.I.

, Sáenz-Lechón

, Osma-Ruíz

and Castellanos-Dominguez

, Dynamic feature extraction: an application to voice pathology detection, Intelligent Automation & Soft Computing 15(4) (2009), 667–82.

Hotelling

, Analysis of a complex of statistical variables into principal components, Journal of Educational Psychology 24(6) (1933), 417.

Cai

, Luo

, Wang

and Yang

, Feature selection in machine learning: A new perspective, Neurocomputing 300 (2018), 70–9.

Jain

and Singh

, Feature selection and classification systems for chronic disease prediction: A review, Egyptian Informatics Journal 19(3) (2018), 179–89.

Muni Kumar

and Manjula

, Role of Big data analytics in rural health care-A step towards svasth Bharath, International Journal of Computer Science and Information Technologies 5(6) (2014), 7172–8.

10.

Eswari

, Sampath

and Lavanya

S.J.

, Predictive methodology for diabetic data analysis in big data, Procedia Computer Science 50 (2015), 203–8.

11.

Preethi

and Asokan

, Modelling LSUTE: PKE schemes for safeguarding electronic healthcare records over cloud communication environment, Wireless Personal Communications 117(4) (2021), 2695–2711.

12.

Hussein

A.S.

, Omar

W.M.

, Li

and Ati

, Efficient chronic disease diagnosis prediction and recommendation system. In 2012 IEEE-EMBS Conference on Biomedical Engineering and Sciences 2012 Dec 17 (pp. 209–214). IEEE.

13.

Guyon

and Elisseeff

, An introduction to variable and feature selection, Journal of Machine Learning Research 3(Mar) (2003), 1157–82.

14.

Tang

, Alelyani

and Liu

, Feature selection for classification: A review. Data classification: Algorithms and applications. 2014:37.

15.

Preethi

and Asokan

, A high secure medical image storing and sharing in cloud environment using hex code cryptography method—secure genius, Journal of Medical Imaging and Health Informatics 9(7) (2019), 1337–1345.

16.

Chandrashekar

and Sahin

, A survey on feature selection methods, Computers & Electrical Engineering 40(1) (2014), 16–28.

17.

Preethi

, Asokan

, Thillaiarasu

and Saravanan

, An effective digit recognition model using enhanced convolutional neural network based chaotic grey wolf optimization, Journal of Intelligent & Fuzzy Systems (Preprint) (2021), 1–11.

18.

Nalband

, Sundar

, Prince

A.A.

and Agarwal

, Feature selection and classification methodology for detecting knee-joint disorders, Computer Methods and Programs in Biomedicine 127 (2016), 94–104.

19.

Hild II

K.E.

, Erdogmus

, Torkkola

and Principe

J.C.

, Feature extraction using information-theoretic learning, IEEE Transactions on Pattern Analysis and Machine Intelligence 28 (2006), 1385–1392.

20.

Sevani and Murthy, Bhabatosh Chanda

C.A.

, Generation of compound features based on feature interaction for classification, Expert Systems with Applications 108 (2018), 61–73.

21.

Preethi

and Asokan

, An attempt to design improved and fool proof safe distribution of personal healthcare records for cloud computing, Mobile Networks and Applications 24(6) (2019), 1755–1762.

22.

Bai

D.P.

and Preethi

, Security Enhancement of Health Information Exchange Based on Cloud Computing System, International Journal of Scientific Engineering and Research 4(10) (2016), 79–82.

23.

Tangri , Kitsios

G.D.

, Inker

L.A.

, Griffith

, Naimark

D.M.

, Walker

and Levey

A.S.

, Risk prediction models for patients with chronic kidney disease: A systematic review, Ann. Internal Med. 158(8) (2013), 596–603.

24.

Sinha and Sinha

, Comparative study of chronic kidney disease prediction using KNN and SVM, Int. J. Eng. Res. Technol. 4(12) (2015), 608–612.

25.

Jena and Kamila

N.K.

, ’Distributed data mining classification algorithms for predicting chronic-kidney-disease, Int. J. Eng. Res. Technol. 4(11) (2015), 110–118.

26.

Lazich and Bakris

G.L.

, Prediction and management of hyperkalemia across the spectrum of chronic kidney disease, Seminars Nephrol. 34(3) (2014), 333–339.

27.

Kolachalama , Singh

, Lin

C.Q.

, Mun

, Belghasem

M.E.

, Henderson

J.M.

, Francis

J.M.

, Salant

D.J.

and Chitalia

V.C.

, Association of pathological fibrosis with renal survival using deep neural networks, Kidney Int. Rep. 3(2) (2018), 464–475.

28.

Sheehan , Mawe

, Cianciolo

R.E.

, Korstanje

and Mahoney

J.M.

, Detection and classification of novel renal histologic phenotypes using deep neural networks, Amer. J. Pathol. 189(9) (2019), 1786–1796.

29.

Kallenberger and Schmidt

, Forecasting the development of acute kidney injury using a recurrent neural network, Cardiovascular Res. 115(14) (2019), e155–e157.

30.

Ren , Fei

, Liang

, Ji

and Cheng

, A hybrid neural network model for predicting kidney disease in hypertension patients based on electronic health records, BMC Med. Information. Decis. Making 19(S2) (2019).

31.

Logeswaran

, Suresh

, Savitha

and Kannan

A.R.

, “Contemplate Study of Contemporary Techniques for HUIM,” in 2018 International Conference on Intelligent Computing and Communication for Smart World (I2C2SW), Dec. 2018, pp. 282–286. DOI: 10.1109/I2C2SW45816.2018.8997506.

32.

Logeswaran

R.K.A. K

, Suresh

, Savitha

, Prasanna Kumar

K.R.

and Ponselvakumar

A.P.

, Data driven diagnosis of cervical cancer using association rule mining with trivial rule expulsion approach, International Journal on Emerging Technologies 11(2) (2020), 110–115.