Report on Machine Learning-Based Prediction for Chronic Kidney Disease

1.0 Introduction of Machine Learning-Based Prediction for Chronic Kidney Disease

For individuals aged 60 or more, kidney disease is recognized as a significant issue. Kidney degeneration is the major cause which decreases the glomerular filtration rate. When it lasts over three months, this problem becomes CKD or chronic kidney disease [1]. In the world, it is ranked as the tenth major cause or reason for death. Aging, diabetes, and hypertension are recognized as leading causes of this disease in addition to some other factors like anemia, disease of coronary artery, and high blood pressure. With the detection of disease in its primitive stages, it is recognized as feasible for saving the function of kidney for the patient's longer survival. The CKD treatment can be facilitated by its diagnosis and it can aid in avoiding costly procedures of treatment like transplants and dialysis.

Lab records and other types of information associated with patients can be analyzed using techniques of machine learning for early CKD detection [23]. With KDD or knowledge discovery in the databases low-level data can be converted into effective and high-level knowledge [2]. Practitioners can be helped in understanding the patterns of CKD for its diagnosis by this transformation.

CKD is analyzed by this study with the use of different techniques of machine learning through a CKD dataset from the data warehouse of machine learning. With the use of Apriori association method, CKD is identified for four-hundred instances of patients with CKD across various classification algorithms such as IBk, J48, naïve Bayes, OneR, and ZeroR. With the normalization and completion of missing data, the pre-processing of dataset is performed. From the dataset, relevant features are chosen for improving accuracy and reducing the time of training for machine learning methods. With the use of different machine learning methods associated with WEKA, experiments are performed for detecting CMD on the basis of dataset of CKD from the UCI instrument [21]. For detection accuracy, results are compared across different techniques of machine learning.

The remaining part of this paper includes: machine learning methods are described by the second section. For CKD detection, previous concepts are presented in the third section while the present work for studying machine learning methods is presented in the fourth section. Results are reported by the fifth section and the paper is concluded in the sixth section with opportunities for future studies.

3.0 Literature Survey of Machine Learning-Based Prediction for Chronic Kidney Disease

Chronic diseases are analyzed and explored by various studies with the use of different methods for diagnosing diseases early. Various techniques of data mining are surveyed by Patil [13] for their accuracy of detection including sequential minimal optimization, k-nearest neighbor, naïve Bayes, radical fundamental function, decision table, ANN, multilayer perception, and logistic regression. In accordance with the dataset type, differences in accuracy levels are indicated by such techniques and for the best result, there is no individual outcome.

The classifier of nave Bayes is utilized by Dulhare and Ayesha [1] with OneR as the selector of attribute for predicting CKD with the use of UCI digital repository’s dataset with twenty-five attributes where thirteen are nominal, one is a class attribute, and eleven are numeric. The attribute number was reduced by them by eighty percent through OneR for an increase of 12.5 percent in detection accuracy.

A clustering technique is employed by Gopika and Vanitha [2] for accurate detection of CKD and decreased time of diagnosis. Techniques of fuzzy k-medoids, k-means, and c-means are utilized by them. In addition to it, accuracy of eighty-seven percent is indicated with the use of fuzzy technique of c-means clustering for a dataset obtained from the machine repository of UCI.

As classifiers for the detection of CKD using a reliable dataset with twenty-four attributes, four-hundred instances, and two classes, k-nearest neighbors, support vector machine, logistic regression, and decision tree are employed by Charleonnan et al. [5]. A CKD dataset is utilized by them from the machine learning repository from UCI. The SVM technique is indicated as a better technique of detection for sensitivity and accuracy of detection by the results.

The failure of kidney function is examined by Ramya and Radha [7] with the use of classification algorithms. In accordance with case severity, kidney disease’s different stages are classified by them with the use of BPNN or back-propagation neural network, radial basic function, and random forest. Different techniques are evaluated by them for different types of performance metrics such as sensitivity, kappa, and specificity while using a dataset obtained from the Coimbatore state for approximately a thousand patients with fifteen attributes. They concluded the radical fundamental function to be most promising classifier with the detection accuracy of 85.3 percent.

The techniques of texture analysis are utilized by Iqbal et al. [9] for analysing the kidney disease’s ultrasound images for distinguishing between kidney disease and normal patients. Mathematical operations are utilized by them like Fourier analysis for calculating the RMS or root mean square, gray-level correlation matrix, average values, and homogeneity. Ultrasound images of approximately thirty-two patients are considered by them and they distinguished between both kidney disease and normal patients with the use of cortex region and RMS values as 0.0049 and 0.3.

Techniques of hybrid classification were employed by Kayaalp et al. [10] for analyzing kidney disease with the use of a dataset from the machine learning repository of UCI with information on approximately four hundred patients. KNN classifier and support vector machine are employed by them. Feature selection is conducted by them with the use of gain and relief ratio algorithm suitable for the dataset feature which is most relevant. It is concluded by them that the algorithm of KNN offers better performance for the specific chosen features compared to other algorithms with respect to contrast matrix, precision, and f-measure.

Boosted classifier and feature selection are utilized by Wibawa et al. [11] for diagnosing CKD and employing AdaBoost for CFS or correlation-based feature selection and ensemble learning. Support vector machine, KNN, and naïve Bayes are utilized by them for detecting and concluding CFS and AdaBoost as reliable and promising classifiers other than naïve Bayes and KNN classifiers in the detection of CKD. 0.98 rate of f-measure, 0.981 rate of accuracy, and 0.98 rate of recall are achieved by them.

CKD is controlled by Wickramasinghe et al. [12] with the use of an eligible diet plan and the recommended plans to a number of patients using their method of classification. A diet plan was recommended by them on the basis of level of blood potassium. Multiclass decision forest, logistic regression, decision jungle, and neural network are utilized by them for achieving the accuracy of 99.17 percent through the use of decision forest algorithm.

It is suggested by previous research that significant insights are provided by machine learning into data and it can aid in classifying data into a number of classes. It is indicated by findings that precise classification results can be produced by machine learning methods if they are utilized together with feature selection methods. Hence, retaining the advantages of classification outcomes for techniques of machine learning, a set of renowned techniques is employed by this study in combination with the techniques of feature selection for classifying normal patients and the ones with kidney disease.