BCD-WERT: a novel approach for breast cancer detection using whale optimization based efficient features and extremely randomized tree algorithm

Introduction

Breast cancer is a life-threatening disease that affects the lives of many women. It is also the second leading cause of deaths occurring due to cancer among females (Meera & Nalini, 2018; Kamel, YaghoubZadeh & Kheirabadi, 2019). Breast cancer annually affects one million people and results in over 400,000 deaths globally (Meera & Nalini, 2018). It is a benign or malignant tumor, caused by uncontrolled growth and division of cells inside the breast (Kamel, YaghoubZadeh & Kheirabadi, 2019). The non-cancerous tumors are called benign which are not life-threatening and can be treated with medicine. The cancerous tumors are known as malignant and if left untreated, the patient can die. These tumors appear as a lump in the breast and can be diagnosed using X-rays. For early detection, the patient must immediately consult with their healthcare provider if such a lump appears in their breast (Meera & Nalini, 2018; Alghunaim & Al-Baity, 2019).

Family history, age, genes, dietary habits, and lifestyle are some of the leading factors causing this disease (Kamel, YaghoubZadeh & Kheirabadi, 2019). Treatments costs are elevated for breast cancer and also, there are severe side effects after long-term usage of medications by patients. Many causalities occur due to the detection of cancer at a stage where it becomes impossible to cure this disease (Kamel, YaghoubZadeh & Kheirabadi, 2019; Alghunaim & Al-Baity, 2019). According to medical professionals, early detection of disease is crucial. Recently, medical practitioners have developed a greater interest in the prediction and detection of breast cancer using machine learning algorithms. Machine learning is a subset of Artificial Intelligence that provides the ability to learn and improve from past experiences without being explicitly programmed (Khourdifi & Bahaj, 2018). In medical discipline, vast and rich information is available about several diseases such as cancer, diabetes, COVID (Sahlol et al., 2020a, 2020b; Dev et al., 2020; Zehra et al., 2020; Khamparia et al., 2020; Gupta, Chakraborty & Gupta, 2019; Chakraborty et al., 2016; Chakraborty, Gupta & Ghosh, 2015). The patterns hidden inside this data can be used to dig in meaningful relationships (Jhaveri & Patel, 2017). Machine learning algorithms are already being incorporated in improving the health, protection and welfare of billions of people (Javed et al., 2020a, 2020b, 2021; Shabbir et al., 2021; Sarwar & Javed, 2019). Many machine learning algorithms such as DT, SVM, RF, Naive Bayes and KNN are being used in the medical field and have provided remarkable results in early-stage disease prediction and diagnosis (Karim et al., 2020; Ak, 2020; Usman Sarwar et al., 2020; Rehman et al., 2020; Javed et al., 2020c).

The virtues of machine learning algorithms are remarkable but there is room for improvement in certain domains such as prediction and detection accuracy, computational complexity, and execution time while using volumes of data (Sagar, Jhaveri & Borrego, 2020; Desai & Jhaveri, 2018). Performance can be improved by careful evaluation and a combination of several optimization techniques available (Kamel, YaghoubZadeh & Kheirabadi, 2019). The performance of these classification algorithms can be improved by using feature selection algorithms such as Grey Wolf Optimization (GWO), Particle Swarm Optimization (PSO) and WOA (Vieira et al., 2013). As the data stored in medical systems is vast and complex, it often contains unnecessary features that may mislead the classification algorithm and negatively affect the results (Chakraborty et al., 2016). Similarly, the relevant features can exclusively optimize the algorithm and correctly predict the class. It is of paramount importance to distinguish relevant features from irrelevant ones and utilize relevant for accurate prediction. Feature selection is used to eliminate the unnecessary data from the dataset and helps improve the accuracy rate of the prediction algorithm by reducing the dimensionality of the data (Kamel, YaghoubZadeh & Kheirabadi, 2019; Alghunaim & Al-Baity, 2019).

In this article, a novel hybrid classification approach named BCD-WERT is adopted that utilizes feature selection and machine learning techniques for breast cancer prediction. BCD-WERT consists of WOA to reduce the high dimensionality of the data and classification is performed using Extremely Randomized Tree (ERT) algorithm. The selected features are given as input to other classification algorithms for breast cancer prediction. These algorithms include SVM, KNN, KSVM, GNB, SGD, DT, LR and RF. BCD-WERT achieves promising classification performance using WOA and ERT in comparison with other classifier and state-of-the-art studies.

The rest of the article is structured as follows: “Literature Review” provides a brief literature review of the several feature selection and data mining techniques used for prediction. In “BCD-WERT Approach”, our proposed hybrid approach is discussed in detail. A brief introduction to the WOA and classifiers is discussed. The experimental results of the proposed approach are provided in “Experimental Analysis and Results”. Discussion of results is presented in the “Discussion” section. “Conclusion and Future Work” presents the conclusion of this research study and highlights dimensions of future research.

Literature review

Breast cancer takes the lives of almost half a million women every year (Meera & Nalini, 2018). It should be detected and cured as early as possible. Several predictions and medical treatment options are adopted by doctors to combat this fatal disease. Early prediction and detection can save many lives. Machine learning algorithms are also being used to predict the disease and have shown promising results in most of the cases. The algorithms proposed so far can be placed in two categories: (a) feature selection based methods and (b) data mining-based methods. The detailed review is as follows:

BCD-WERT Approach

In this study, a classification based method named BCD-WERT is proposed for breast cancer prediction. BCD-WERT uses WOA based efficient features and well-known classification algorithms. BCD-WERT consists of four main phases. In phase I, the dataset is collected for experiments. The Wisconsin breast cancer diagnostic dataset from Kaggle online dataset library is used in this study. In phase II, data is preprocessed and cleaned accurately for prediction. In phase III, the WOA is used for feature selection. In phase IV, selected features are used for the training of the classification models. Several classification methods are used for the classification of testing data. For this study, the WOA and all other classification models are implemented using Python. Figure 1 presents the working of the proposed approach and a detailed Algorithm 1 is presented below.

Experimental analysis and results

For experimentation, the Wisconsin breast cancer dataset (Asri et al., 2016) is used for the detection of breast cancer. The dataset is split into 75% for training and 25% for testing. To analyze the performance of the proposed classification model, the performance measures such as accuracy, f-score, recall, and precision are used. The best feature is predicted using WOA and then classification models perform prediction. The proposed model reduces the number of features and is capable of handling large data for cancer prediction with better accuracy. BCD-WERT achieves the highest accuracy of 99.30% with Extremely Randomized Tree (ERT) classifier. The performance of other classifiers can be observed in Table 2.

• Support Vector Machine: SVM classified the data with 98.60% accuracy making use of selected features extracted through WOA. The precision of both classes 0 and 1 is 1.00 and 0.98 respectively. Precision shows that the number of correctly identified positive results in both classes is more than the incorrect ones. F-score of the classification model for 0 and 1 class in the SVM case is 0.98 and 0.99 respectively and the recall for both classes is 0.96 and 1.00 respectively. These results depict that SVM can be used to predict breast cancer and early diagnosis is possible for most cases. Receiver Operating Characteristics (ROC) curve and confusion matrix of SVM classification is presented in Fig. 2.

• Extremely Randomized Tree: ERT outperforms all other classifiers used in our experiments and achieves an accuracy of 99.30%. This is achieved since this classifier aggregates results of multiple co-related DTs and uses a random sample for this purpose. The accuracy achieved shows that this classifier can be effectively used to predict this fatal disease. The f-score of both 0 and 1 class, in this case, is 1.00 and 0.99 and the recall for 0 and 1 is 0.98 and 1.00 respectively. Precision shows that correctly identified positive results are enough and the value for both classes is 0.99. ROC curve and confusion matrix of ERT classification is presented in Fig. 3.

• Logistic Regression: LR classification is used to estimate the perimeters of the logistic model and accuracy achieved using LR classifier is 97.01%. The precision measure of regression for 0 and 1 class is 1.00 and 0.96 respectively and the f-score achieved is 0.96 and 0.98 respectively. The recall of the classifier is 0.98 and 0.96 respectively. ROC curve and confusion matrix of LR classifier can be seen in Fig. 4.

• Random Forest: RF gives high accuracy with large datasets. The accuracy after applying this classifier on the Wisconsin breast cancer dataset is 98.60% and the f-score calculated for 0 and 1 class is 0.98 and 0.99 respectively. The precision measure for RF is 0.96 and 1.00 for 0 and 1 class respectively. The recall of this classifier for classes 0 and 1 is 1.00 and 0.98 respectively. These results show that RF performs well on a large dataset. ROC curve and confusion matrix of RF classification is presented in Fig. 5.

• Kernel Support Vector Machine: KSVM operates in high dimensions and performs pattern analysis. The accuracy achieved by KSVM in the proposed solution is 94.41% and precision for class 0 and 1 is 0.98 and 0.93 respectively. The F-score for KSVM in both 0 and 1 class is 0.92 and 0.96 respectively. Recall for the KSVM classifier obtained is 0.87 and 0.99 for both 0 and 1 class respectively. ROC curve and confusion matrix of KSVM classification is presented in Fig. 6.

• Gaussian Naive Bayes: GNB is used to find strong independence between the features of the dataset and it performs well when applied to the selected features of a huge dataset. The accuracy achieved by the classifier is 94.40% and the precision measurement of the 0 and 1 class is 0.91 and 0.97 respectively. The F-score achieved by using GNB is 0.93 and 0.96 for both 0 and 1 class respectively. These results depict that this classifier can be used for the prediction of cancer but other classifiers outperformed GNB in terms of accuracy. ROC curve and confusion matrix of GNB classification is presented in Fig. 7.

• Stochastic Gradient Descent: SGD performs well on large scale dataset but it is sensitive when it comes to feature scaling. The accuracy achieved by the classifier is 93.00% and the F-score for 0 and 1 class is 0.90 and 0.95 respectively. The precision and recall achieved for 0 and 1 class are 0.98, 0.91 and 0.83, 0.99 respectively. ROC curve and confusion matrix of SGD based classification can be observed in Fig. 8.

• K-Nearest Neighbour: KNN relies on distances for classification and training and here data normalization can improve accuracy. The accuracy achieved by the KNN used in the proposed solution is 93.20%. F-score and precision of class 0 and 1 is 0.90, 0.95 and 0.92, 0.93 respectively.The results obtained by using K-Nearest Neighbour are adequate when there is no other classifier but it performed worse than the other classifiers used because of the value of the k selected on a random basis. The recall of the proposed solution is 0.89 and 0.96 for classes 0 and 1 respectively. ROC curve and confusion matrix of kNN classifier is presented in Fig. 9.

• Decision Trees: DT is used to split data according to a certain parameter. The accuracy achieved by this classifier is 97.20% and the F-score and precision of class 0 and 1 are 0.96, 0.98 and 0.95 and 0.99 respectively. The recall obtained is 0.98 and 0.97 for 0 and 1 class respectively. ROC curve and conclusion matrix of DT based classification is presented in Fig. 10.

Discussion

There exist several randomization methods for trees and classification but building completely random trees are still far away. Extra Randomized Tree algorithm works solely on random behavior and selects the cut points for the tree. In the end, a randomized tree is built and then prediction and classification are done using that tree. The proposed approach is capable of accurately predicting whether a patient has breast cancer or not, even when limited data is available as it only chooses the best feature for classification. Medical practitioners can easily use this model to check if a patient is at risk of getting breast cancer. Nowadays, medical facilities are getting expensive so this model can help in the early diagnosis of cancer in patients (Sivakami & Saraswathi, 2015). Table 2 shows the results obtained from the proposed approach and the ERT classifier outperforms all other classifiers used for breast cancer prediction. The accuracy achieved through SVM, ERT, DT, LR and RF is higher as compared to other algorithms used in this study. These can be used to classify and predict breast cancer to avoid the spreading of cancerous cells in the body. If the disease is diagnosed timely, many patients can be saved from facing dire conditions in later stages and the treatment of this disease can be started on time. All classifiers used in this study performed well except SGD and KNN. The reason behind this is because the KNN classifier works on distance measures, the value of k also has an immense effect on the accuracy of the model. SDG performance lacks because of the usage of the same learning rate for all parameters. All the other classifiers were able to classify the dataset with a high accuracy rate. ROC and confusion matrix of each classifier shows the performance of the classifier and ERT proves to be the best classifier among all nine classifiers used in this study. Table 3 shows the result summary and comparison with the existing solutions. The ROC curve and confusion matrix of each classifier is given.

Conclusion and future work

In this paper, a novel approach named BCD-WERT is proposed that utilizes WOA and Extra Randomized Tree (ERT) algorithm for the detection of breast cancer. WOA based feature selection is done to extract optimal features from the dataset and to eliminate any unnecessary details. This is given as input to the ERT classifier and other algorithms. DT, KNN, SGD, RF, LR, KSVM, GNB and SVM are also used for classification and performance is compared with BCD-WERT. The results showed that BCD-WERT achieved the highest accuracy rate of 99.03% by using the WOA and ERT classifier. In the future, we intend to develop a software application to enable user predict if the cancer is benign or malicious. This application would benefit the society and will help the medical community to detect cancer at early stages. Machine learning and deep learning algorithms can be applied to the dataset to achieve better results and can be used in other areas of medical discipline.

Supplemental Information