Prediction of Chronic Kidney Disease with Machine Learning Models and Feature Analysis Using SHAP

Chronic kidney disease is one of the concerning diseases faced by many people around the world. It mainly causes the decrement of kidney functioning in filtering the wastes. The kidney's main job is to remove waste and extra fluid from the blood circulation, which is then expelled as urine. The early stages of CKD may be asymptomatic, meaning that people may not experience any symptoms until the condition has progressed. Common symptoms of advanced CKD include fatigue, weakness, loss of appetite, muscle cramps, and swelling in the legs and ankles [1].

A person with high blood pressure, diabetes may get affected with CKD [2]. CKD can lead to various complications including anemia, bone disease, nerve damage, and an increased risk of cardiovascular disease. It may lead to end-stage kidney disease which often leads to dialysis or kidney transplantation. These treatments are very expensive. Hence, the disease needs to be detected as early as possible to prevent or delay the onset of these complications.

Since the role of kidney is to filter blood and get rid of wastes through urine, it is obvious that we can get initial and primary information from Blood and urine test to detect CKD. Generally, it is expensive and time taking to take tests and consulting doctor. Moreover, every place may not have efficient doctors or test Labs. Currently, data mining and artificial intelligence are ruling the world. Huge data is available online related to various sectors. Grabbing this advantage into healthcare, CKD detection can be made automated i.e., without actually consulting doctor (for primary validation).

Digitalizing the CKD diagnosis process will be handier and available to every person irrespective of location which is possible using Machine Learning (ML) Technology.

Machine Learning algorithms are fit for analyzing vast amount of patient data and detecting patterns that may not be apparent to a human clinician. This ability enables the identification of patients at risk of developing kidney disease at the early stage, even before symptoms emerge. By intervening early, clinicians may potentially slow or even prevent the progression of kidney disease.

The need to use ML in kidney disease prediction is driven by the potential to improve early detection, personalize treatment plans, improve accuracy, and reduce healthcare costs [3]. ML can provide a valuable tool for clinicians to improve patient outcomes and manage the growing burden of kidney disease [4].

Transparency is crucial for gaining trust in medical domain projects, particularly when black box algorithms are utilized, which can obscure the internal workings of a model. To address this issue and ensure code interpretability, Shapely values can be plotted using various types of visualization techniques to display the contribution of each feature to a particular classification result [5]. By providing this level of detail, users can better understand the reasoning behind a model's decision-making process. Furthermore, these results can be shared with medical professionals, enabling them to make informed decisions and develop tailored treatment plans for patients.

This work mainly aims to create easier way to diagnose CKD through corresponding clinical details using Machine Learning. Logistic Regression, Random Forest, Support Vector Machine and Voting Classifier algorithms are going to be used for classifying the patients with CKD and non-CKD. This makes the initial test as cost effective for every class of people. This work also uses model interpretability technique – SHAP (SHapley Additive exPlanations) for analysing the feature contribution.

3.1 Preprocessing

The dataset is containing missing values. It is important to pre-process the data before giving to the models. Since selected dataset is small in size it is not suitable to merely remove the records with missing values from the dataset. Instead, the missing values are replaced by the mode frequency of their corresponding columns (used mode imputation) [9].

Dataset contains categorical columns which are transformed into numerical. The values such as ‘good’, ‘present’, ‘yes’, ‘normal’ are replaced with 1 and values ‘bad’, ‘not present’, ‘no’, ‘abnormal’ are replaced with 0. Later, dataset in transformed where scaling is applied to normalize all the values [10].

3.2 Machine learning models

3.2.1 Logistic Regression

A common method of supervised learning used to categorize binary situations is Logistic Regression (LR). Using one or more independent variables, it calculates the likelihood of a binary result for the dependent variable. In LR, output is transformed using Sigmoid function, which converts any real-valued input to a number between 0 and 1. A threshold is then applied to classify instances as positive or negative. LR can handle relationships between the dependent and independent variables that are both linear and nonlinear [11, 12].

3.2.2 Support Vector Machine

Support Vector Classifier (SVC) is a supervised learning technique that uses a hyperplane to split the data into two classes. The algorithm seeks to determine a decision boundary in order to maximize the margin and minimize classification errors between the two classes. This is achieved by identifying support vectors. These are the closest data points to the decision boundary. Once the hyperplane is identified, new data points can be classified by determining which side of the hyperplane they belong to. The kernel function in SVC can be used to deal with non-linearly separable data by projecting the input to a higher-dimensional space [13, 14].

3.2.3 Random Forest

It is a kind of ensemble learning algorithm that combines multiple decision trees to improve accuracy and prevents overfitting. The algorithm randomly selects a subset of features and data samples from the training set to create each decision tree. During the classification stage, the algorithm aggregates the predictions of all decision trees to make a final classification decision. It has the ability to handle large datasets and high-dimensional data. It is robust to noise and outliers. It is a popular choice among Machine Learning algorithms due to its versatility [15, 16].

3.2.4 Voting Classifier

Voting Classifier is a different type of ensemble learning algorithm that combines multiple individual classifiers with different algorithms and hyperparameters. During classification, each individual classifier makes its prediction, and the Voting Classifier aggregates these predictions to make a final decision. This can be done either by hard voting or soft voting. In hard voting, the class with the highest number of votes determines the output. In soft voting, the final output is based on the class with the highest average probability calculated from the predicted likelihoods of each class by all the individual algorithms. Here, LR, SVM, and Decision Tree Classifier are used in Voting Classifier [17].

The work flow of this proposed work goes as shown in Figure 1.

image006.jpg

Figure 1. Methodology

Performance evaluation

Performance evaluators in ML are metrics or measures used to evaluate the accuracy and potency of a model in making predictions on new data. They help to compare different models and identify areas for improvement [18].

Confusion matrix:

A 2×2 confusion matrix is a standard matrix used to assess the performance of binary classification models. It consists of four cells that represent the four possible outcomes of a binary classification problem, as shown in Figure 2.

image008.png

Figure 2. Confusion matrix (for Voting Classifier)

Eqns. (1), (2), (3) and (4) are the formulae for the metrics calculation.

Accuracy: The proportion of percentage of correctly predicted labels to the total number of predictions.

Accuracy $=\frac{T P+T N}{T P+F P+T N+F N} \times 100$                     (1)

Precision: The proportion of correctly predicted positives to the total predicted positives.

Precision $=\frac{T P}{(T P+F P)} \times 100$                     (2)

Recall: The proportion of correctly predicted positives to the actual total positives.

Recall $=\frac{T P}{T P+F N}$                     (3)

F1-score: The harmonic mean of recall and precision.

$F 1-$ Score $=2 \times \frac{\text { Precision } * \text { sensitivity }}{\text { Precision }+ \text { sensitivity }} \times 100$                     (4)

SHAP

SHAP (SHapley Additive exPlanations) [19] is a model-agnostic technique for describing the predictions of ML models. It is based on the notion of Shapley values, where each feature is assigned with a value in a prediction based on how much it contributed to the outcome. In the context of Machine Learning, these values measure the impact of each feature on the model's prediction for a particular data point. They provide a detailed understanding of how the model arrived at its decision by quantifying the influence of each feature on the output. By utilizing SHAP values, it is possible to identify the most influential features in a prediction and visualize how they contribute to the final result. This can help to improve model interpretability, identify biases, and facilitate more effective decision-making in Machine Learning applications [20].

By providing feature-level explanations of model predictions, SHAP can help healthcare professionals understand how a model is making its decisions, which can lead to better trust and adoption of these models. It ensures that models are fair and transparent, which is essential for ethical use in the medical domain [21-23].

This paper has explored a Machine Learning approach for predicting Chronic Kidney Disease (CKD) using a dataset from the UCI Machine Learning Repository. Data preprocessing steps, including data cleaning, feature selection, and feature scaling are performed. The models Logistic Regression (LR), Random Forest (RF), Voting Classifier and Support Vector Machine (SVM) are trained and evaluated using various metrics to estimate their performance. SHAP values cannot be used for causal inference. This is the process of finding the true causes of an event/target. SHAP values tell us how each model feature has contributed to a prediction. They do not tell us how the features contributed to the target variable.

Overall, the results demonstrate that Machine Learning can be a powerful tool for predicting CKD, and that the Voting Classifier and Random Forest performed well and Random Forest model with SHAP values can provide useful insights into the underlying patterns and relationships in the data.