Abstract
Cancer is a disease caused by uncontrollable cell growth. The disease is a constant subject of concern due to unavailability of treatment at a severe level. Patients who have suffered from the disease have the chance of getting saved if this fatal illness is identified in the beginning stage. The survival chance will be very low if it is detected in the final stage of cancer. As the patients could not survive in their last stage, to cure their disease, an early diagnosis is a key issue and is vital. For the classification of cancer, Gaussian Naïve Bayes is implemented in this work. By exerting it on two datasets, the algorithm is tested, in which the Wisconsin Breast Cancer Dataset (WBCD) is considered as earliest one and the next one is the Lung Cancer Dataset. The assessment result of the suggested algorithm attained 90% accuracy in the prediction of lung cancer, and in predicting breast cancer, the accuracy is 98%.
1. Introduction
Classification is a task of searching or mining data that classifies items in an accumulation to the target classes [1]. In data, where the classes are prebuilt, the classification predicts each instance’s target class and it is a supervised technique of learning [2]. The objective of classification is, that for every status in the data, the target class should be predicted accurately. Classification helps in the discovery of various models of concealed data that have been not transparent to the data annotator [3]. Medical diagnosis, detection of faults in industry, and recognition of patterns and images are some examples of applications of classification. It is the main side that helps us to select and make decisions [4] between two or more groups. The two parameters which are used to find the classification model’s coherence are accuracy and interpretability. The disease of cancer in many cases brings about the death of the majority of people who suffer from it [5]. Cancer is caused by the unusual extension of cells in any parts or tissues of the body. It may take place in any part of the body. We might be thinking that all types of tumors are not cancerous but the fact is, yes. Since it is basically unnecessary, some tumor types do not unfold to all parts of the body. The different indicators of cancer include tumor, an abnormal type of bleeding, cough for a considerably long time, and loss of weight. Earlier recognition [6] is the most productive way to abate the deaths of cancer. Without going for surgical operations, early detection needs a well precise and dependable procedure of detection that permits doctors to separate benignant lumps from cancerous ones [7]. Altogether, there are more or less than 100 kinds of cancers that affect the body of human beings. Some kinds of cancers are as follows:(a)Of the two datasets, lung cancer is the anomalous cell’s unusual growth that undertakes in a single lung or both of the lungs usually in the form of cells lining the passages of air [8]. The healthy tissues can be destructed by the growth of abnormal cells; they divide quickly and the formation of tumor occurs. Basically, cancer in the lungs is divided into two. They are nonsmall cell lung carcinoma (NSCLC) and small cell lung carcinoma (SCLC). Furthermore, NSCLC is broadly divided into adenocarcinoma, large cell carcinoma, and squamous cell carcinoma. SCLC is the second type. There are also various histologic groupings in SCLC which include large cell carcinoma, combined small cells, pure small cells, and also mixed small cells. The more attacker than NSCLC is deadlier than SCLC [9](b)The other dataset describes that cancer which affects females worldwide is breast cancer [10]. It remains to be one of the most popular cancers in the world. In 2017, from all of the cancer status, the estimation seems to be 25% that detected amidst 1.67 million statuses of cancer were breast cancer. Specifically, compared to the developed nations, the women from less developed nations have a diminutive number of cancer statuses, like 794000 in advanced nations and 883000 cancer statuses in less advanced nations [11]. In comparison, the breast cancer rate in the United Kingdom (UK) is higher than the case of India where the rate is 28.8 per 100000, while in the case of UK, the rate is 95/100000, but the rate of death equals (12.7 vs. 17.1/100000) that of the UK.
There are several techniques of classification in literature, which are familiar with the detection of cancer like the genetic algorithm, C4.5, and using support vector machines (SVMs) [12]. Gaussian Naïve Bayes is implemented in this study to classify cancer that aids doctors to diagnose this deadly infection in the early stage and to keep the patient’s life [13].
2. Related Works
Data mining techniques are used by several researchers for the detection of various diseases. Jyothy Sony deployed classifiers such as K-nearest neighbors [14] and Naïve Bayes. Classifiers have been used to datasets of Wisconsin Prognostic Breast Cancer (WPBC) [15] in Naïve Bayes of Diana Dumitru’s research, concerning a binary class: instance’s nonrecurrent events and 47 instances’ recurrent events and also several 198 patients. Mohd Fauzi et al. presented different classification techniques’ comparisons such as radial basis, Bayes network, and pruned tree. On a large dataset, for knowledge analysis, the K-nearest neighbors algorithm uses Waikato to Environment. Breast cancer is the data used in their investigation. Alazab et al. [16] analyzed the supervised learning algorithm’s performance such as SVM Gaussian RBF Kernel, RBF neural networks, simple CART, decision tree, and Naïve Bayes. These algorithms are accustomed to classifying the datasets of breast cancer [17]. Shandilya and Chandankhede designed the system of an intelligent and effective prediction of heart disease using the associative classifier’s weighted concept [18]. As compared to the other already existing associative classifiers, it gives better accuracy. The accuracy was found to be 81.51% for weighted associative classifiers (WAC). According to their capability in predicting, WAC assigns different weights to different attributes. With regards to accuracy [19], experiments are being conducted on a dataset of benchmark to assess the performance [20].
3. Naïve Bayes
It is an algorithm of effortless learning that utilizes the rule of Bayes in along with a high premise or assumption whose characteristics are depending on the independence provided by the class. Though Naïve Bayes brings the accuracy of competitive classification again and again, while assuming this independently often invaded in practice [21]. Practically, the combination of computational efficiency and many other advantageous aspects may lead to naïve Bayes being largely applied. Evaluate the posterior probability value of , each of the class of y, with an object x. Using the details inside the sample data, Naïve Bayes contributes an apparatus. If such evaluations are done once, they can be used for the classifications or any other application of decision support [22].
The following are the features of Naïve Bayes:(i)Low variance As the search is not utilized by Gaussian Naïve Bayes, it contains variance at a low value, despite the cost of bias being high(ii)Incremental learning In general, the Gaussian Naïve Bayes functions from probabilities of the lower-order estimates obtained from that of the training data. As new training data are obtained, these can be updated readily.(iii)Computational efficiency The training time equals the attributes and training examples, while the value of classification time equals the attributes and it is unaffected by the attributes number.(iv)Robustness in the face of absent value If one of the attribute values is missing and information from another attribute is still used, then it results in the sophisticated deterioration in performance, as for all predictions, all attributes are always used by the naïve Bayes. Due to its probabilistic framework, it is extremely thoughtless of the value of missing attributes in the data deployed in the training process.(v)Robustness in facing noise
For all predictions, Gaussian Naïve Bayes uses the whole attributes. Hence, it is verbally thoughtless to noise in the examples which is needed to be assorted and also for noising in the training data as it uses probabilities.
4. Naive Bayes Classifier
Learning can be greatly simplified by the Naïve Bayes classifier by supposing that features are independent given class [23]. Although, the assumptions of independence are poor in general. Practically, with a more sophisticated classifier, Naive Bayes often competes effectively. The feature vector described that to a given example, Bayesian classifiers allocate the most likely class, i.e.,where C is the classifier, and is a feature vector [24].
In practice, the resulting classifier known as Naïve Bayes is very successful despite this unrealistic assumption, by competing again and again with much more sophisticated techniques. The model of Naïve Bayes is a simplified version of Bayesian probability. The operations of the Naïve Bayes classifier are done on a sturdy independence assumption. This means that one attribute of the probability of one does not have any effect on the probability of the rest of the attributes [25]
For the feature selection, rank search and Cfs subset evaluator are used. Therefore, the adoption of another technique of rank search has been tried. 89 attributes are selected, and chi-square feature selection has been adopted this time. 89 attributes have been inputted this time in preference to the previously selected 75 attributes. After using feature selection of chi-square, the result has been improved, 0.1% of improved accuracy. Even though there is an insignificant improvement, it is certified that to achieve better classification results, feature selection and preprocessing are really useful, as given in Table 1. It is believed that under different situations, to fulfill different classification results, different searching techniques can help. To generalize the best solution, there are needs to take many trials and time. However, due to the constraints of time, this study concludes only by using feature selection and preprocessing which can attain better classification outcomes [26, 27].
5. Cancer Classification
Clinical prediction of results in cancer is generally accomplished by the evaluation of histopathology of samples of the tissue procured during primary tumor surgical removal. The characteristics of histopathology include tissue integrity, the size of the tumor, histological grade, a typical cell morphology, irregular expression of genetic markers and protein, malignant transformation’s evidence, proliferation and senescence, invasion depth, invasive margin’s characteristics (IM), and the extent of vascularization [28]. By the system of classification, the progression of cancer’s evaluation is carried out longitudinally and then put in to estimate patient prognosis. It is important to begin the assimilation of the immunocore as a cancer classification component as it is considered the imperative role of the host immune signature to control the progression of tumor and a prognostic tool. The Gaussian Naive Bayes algorithm is shown in Algorithm 1. There are two advantages to this strategy.(1)It seems to be the strapping symptom in the case of overall survival [29, 30] (OS) as well as disease-free survival (DFS), especially in the first stage of cancer(2)For novel therapeutic approaches, it could suggest potential goals including immunotherapy; concerned with procedure of prognostic and diagnostic analysis of tumors, the technologies of ongoing immunohistochemical allow the application of such assessments by empirical labs.
6. Proposed Method
|
We can see the scheduled work’s main phases in this section (classification of cancer by using Naïve Bayes) as shown in Figure 1.
6.1. Data Acquisition Stage
From cancer, the classification proposed was executed in two sets.
For classification, from the machine learning bin of UCI, the Wisconsin Breast Cancer is quoted as a standard dataset where the database contains 699 instances and 11 attributes are the first type. The dataset of Lung Cancer which is procured from the original website where the database contains 23 attributes and 1000 instances of data (expecting ID number and class label) is the second type. It describes gender, age, pollution of air, allergy to dust, use of alcohol, genetic risk, occupational hazards, obesity, balanced diet, chronic lung disease, passive smoker, smoking, coughing of blood, chest pain, fatigue, shortness of breath, weight loss, snoring, frequent cold, swallowing difficulty, dry cough, wheezing, and clubbing of fingernails. The attributes of the Breast Cancer dataset are given in Table 2.
6.2. Data Preprocessing
In data mining, data preprocessing acts as the initial step. Before exerting data preprocessing for any type of classification, the database that is made useful in research must be preprocessed. This dataset may hold on data and leads up to deceiving results. Therefore, the data’s quality should be improved before working on an evaluation. Data normalization is the data preprocessing done for this work. The technique which was used to sniff out the attribute values that are worthy of being zero and are not effectual in this procedure of classification is the z-score normalization technique.
7. Results and Discussion
7.1. Accuracy
Accuracy discloses, in general, how the model works and whether a model is being trained properly and correctly. However, it does not give specific details on how it tried to the problem. As your primary success, the demerit in using machine learning (ML) accuracy is that, when there is a significant class divide, it fails inadequately. Accuracy may be false for the unbalanced datasets. For evaluating classification models, accuracy is one metric. It has the following definitions formally:
Accuracy is calculated with the positives and negatives for binary classification as follows:
where TP is the value of true +ive; FP is the value of False +ive; TN is the value of true −ive; and FN is the value of false −ive.
7.2. Precision
When cost of false positives (FP) is high, then the precision helps. So, let us assume that the issue involves skin cancer detection. Many of the patients will tell that they have melanoma if we have a very low precision model and that may include some false detection. After being assaulted with false alarms, those who detect the results will ignore them when the false positives are too high.
7.3. TPR
TPR (true positive rate), which is also called recall, is metric related to the classification model. It is calculated as
The probability that an actual positive will test positive is true positive rate.
7.4. DSC
DSC (dice similarity coefficient) is also known as Sorensen Dice Index. To measure the similarity of two samples, DSC is used and it is a statistic matrix.
By using the definition of the value of true +ive (TP), false −ive (FN), and false +ive (FP), when applied to Boolean data, it can be written in the form of
For the implementation of algorithms, the Visual Basic Net.2013 is viable [29]. We have executed assessments on datasets of breast cancer which consists of 699 instances, of which 241 instances are huge and 458 instances are benignant and also applied on datasets of lung cancer containing thousand instances, of which more than 300 records of the file are of lower level, 332 records of the file are of the medium level, and 365 high-level records for the effectiveness of evaluation of the proposed algorithm. For learning and training, each dataset is split into 70% and for testing, 30%. For the first dataset, it is benign or malignant to segregate the data. For the second dataset, categorize the data irrespective of its level such as low, medium, or high level. Three metrics have been calculated, for assessing the execution system proposed. These metrics and their formulas are
where TP indicates true +ive and TN indicates true −ive and FN and FP indicated false +ive and false −ive, respectively.
A confusion matrix is a basically a table of data that holds the details about predicted and actual classification systems attained by a system of classification. Data used for the evaluating the execution of such a system are given in Table 3.
Notwithstanding, to compute specificity, accuracy, and sensitivity for both datasets, these values of metrics were used, as given in Tables 4 and 5, which indicate the matrix of confusion for the preprocessed system. Table 6 provides the result of the Naïve Bayes classifier’s performance evaluation of Gaussian Naïve Bayes.
In cancer classification, empirical effects display that the algorithm was the best as shown in the other previously existing algorithm comparison in Table 7 and Figure 2.
8. Conclusion
To identify the disorganized value in the classification for normalization, the proposed work presents an effective approach, which deserves to be zero. For the prediction of cancer based upon Gaussian distribution, the algorithm of Naïve Bayes is used. Among the two types of cancer named lung cancer as well as breast cancer, the model was implemented, and to list the patient’s details, an interface has been intended, which predicts the class of cancer. Therefore, high classification performance has been achieved by the proposed work. In this study, the author’s is early detection of cancer using the Naïve Bayes algorithm. The survey presented in this study shows different cancer classifications and found that Naïve Bayes is an algorithm of simple learning and learned about Naïve Bayes classifiers. This research helps in the success of the project.
Data Availability
The data used to support the findings of this study are available from the corresponding author upon request.
Conflicts of Interest
The authors declare that they have no conflicts of interest.