¹Microsoft , Sr Technical Support Enginner
²Microsoft , Sr Technical Support Engineer
³Code Ace Solutions Inc, Software Engineer
⁴JP Morgan Chase, Lead Software Engineer
⁵Deloitte Consulting LLP, Senior Solution Specialist
⁶North Star Group Inc, Software Engineer

*Address for Correspondence:Purna Chandra Rao Chinta, Microsoft, Sr Technical Support Enginner Email Id: chpurnachandrarao@gmail.com

Submission: 04 January, 2025 Accepted:31 January, 2025 Published:03 February, 2025

Copyright: © 2025 Chinta PCR, et al. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

Keywords: World Health Organization (WHO); Breast Cancer; Tumour; Machine Learning; Healthcare; Disease Diagnosis; Feedforward neural network (FFNN); Random Forest (RF); Decision Tree (DT); Convolution Neural Network (CNN)

At now, breast cancer ranks second among women in terms of cancer-related deaths, making it a major epidemiological issue. The illness is not caught early enough, and half of the one million women diagnosed with breast cancer annually die from the condition. This research aims to predict the occurrence of breast cancer using various ML algorithms, including Feed forward Neural Network, Random Forest, and Decision Tree, with the goal of reducing the risk of death from this disease, which is a second most common cause of death among women globally. This research uses the Breast Cancer Wisconsin (Diagnostic) dataset to assess ML models that may diagnose breast cancer. The FNN model outperformed RF and DT, achieving the best overall performance with a precision, recall, and accuracy of 97.18%. These results highlight the FNN’s robustness in minimising false positives and maximising true positives, making it a reliable tool for breast cancer diagnosis. To further enhance the accuracy of feature extraction and classification, future research may look at incorporating stronger deep learning models such transformer architectures and Convolution Neural Networks (CNNs). The model’s generalisability and clinical usefulness might be further validated by using bigger and more varied datasets.

In terms of mortality rates, cancer ranks high among both sexes worldwide. The most common malignancies are breast cancer, which kills more women than any other illness and affects more women than any other disease in the world. Breast cancer may be detected early, which could lead to a survival rate of up to 80%, according to the WHO [1]. There are almost 1.7 million new instances of breast cancer identified each year, with 500,000 people losing their lives to the condition. Unfortunately, these figures might rise in the years to come [2,3]. Dense breast tissue, a personal or family history of breast cancer, an older maternal age, the use of certain medications or procedures during pregnancy, drinking alcohol, and other behavioral variables are all potential risk factors for this kind of cancer [4]. The impact of certain factors is substantial, while that of others is rather little. Being a woman and getting older are unchangeable facts, but may lessen our risk of breast cancer by living a healthy lifestyle.

There are three main ways to detect breast cancer: a physical exam, a mammogram, or a biopsy. Professional radiologists are required to interpret the results of these diagnostic procedures; nonetheless, mammography is by far the most prevalent [5]. The problem with having several readings of the same mammography is that various radiologists get different conclusions. There is a 65% to 78% accuracy rate for mammography. The malignant nature of a tumour found by mammography may be determined by doing a biopsy [6]. Although the accuracy rate of a biopsy is almost 100%, the procedure is nonetheless invasive, expensive, time-consuming, and unpleasant [7]. These issues make it more challenging for clinicians to diagnose benign or malignant tumours. Because of these factors, ML techniques have the potential to greatly impact the diagnostic process.

The application of AI techniques for the early diagnosis of breast cancer has recently increased. Learning theory is one kind of AI. For the most part, healthcare organisations have used ML and DL algorithms for breast cancer diagnosis [8]. The diagnostic accuracy of a patient utilised to be entirely dependent on the knowledge and skill of the doctor [9]. The accumulation of a physician’s expertise is the result of years of closely observing patients’ symptoms [10]. However, the accuracy is unreliable. It is now simpler to collect and store data because of the development of computer tools. Thus, the field of intelligent healthcare systems is dependable and beneficial. These technologies may assist doctors in diagnosing patients by providing them with relevant and reliable standards. Individuals might also benefit from these developments in terms of future health planning. This is how ML can take over the laborious physical tasks that healthcare workers face every day [11,12].

This project aims to explore the use of ML algorithms on the Breast Cancer Wisconsin (Diagnostic) dataset to assess predictive analytics’ potential in disease diagnosis. By determining which model, out of many options including FNN, RF, and Decision Tree, performs the best, the study significantly advances medical diagnosis. The key contributions are:

Collect the Wisconsin Diagnostic Breast Cancer dataset for breast cancer detection. Applied essential data preprocessing steps, including the removal of duplicates, handling missing values, enhance data quality and model performance.

Applied standardisation to scale the features of the dataset, transforming them to have a mean of zero and unit variance. Apply ML models like FNN, RF, DT for breast cancer detection. Evaluated model performance employing accuracy, precision, recall, F1-score, and AUC, focusing on a comprehensive understanding of predictive capabilities, especially for imbalanced datasets.

Presented below is the outline of the paper: Section II finds research gaps and evaluates pertinent literature. Section III details the methodology, including data collection and the machine learning models used. Section IV presents the results of the experiments and the analysis of the model’s performance. Section V wraps up the report by reviewing the results and offering suggestions for further study.

They summarise the research on breast cancer prediction and categorisation in this section. Classification methods were the primary emphasis of the literature studied. Some reviews are:

In this study, Khuriwal and Mishra (2018) proposed using the Wisconsin Breast Cancer database in an adaptive ensemble voting method for breast cancer diagnosis. This study aims to examine and explain how logistic and ANN algorithms, in conjunction with ensemble ML algorithms, produce improved results for breast cancer diagnosis, even when the number of variables is decreased. Wisconsin Diagnosis Breast Cancer was the dataset used in this research. When contrasted with similar literature. The Artificial Neural Network (ANN) technique achieved a 98.50% accuracy rate while utilising the logistic algorithm, according to an alternate ML methodology [13].

In this study, Gecer et al. (2018) provide a method for creating five diagnostic categories from breast biopsy whole slide images (WSI). A saliency detector using four fully convolutional networks trained with data extracted from pathologists’ screening records is an integral part of the WSI diagnosis process. After that, this detector will locate diagnostically important regions using multi-scale methods. Then, image patches are classified using a convolutional network based on whether they are invasive cancer, ductal carcinoma in situ, atypical ductal hyperplasia, proliferative changes, or non-proliferative. The network is trained using reference samples collected from consensus. At last, the saliency and classification maps are combined to label pixels and classify slides, respectively. Both the saliency detector and classifier networks outperformed rival algorithms in experiments, including 240 WSI. There was no significant difference between the 45 pathologists’ opinions and a five-class slide-level accuracy55%. Breast cancer diagnostic visualisations using the learnt representations are also offered [14].

In this study, Chen et al. (2017), optimise techniques for ML to accurately forecast the onset of chronic diseases in populations prone to such outbreaks. A chronic illness of the brain, cerebral infarction, is the subject of our experiments. They present a novel multimodal illness risk prediction method that utilises hospital structured and unstructured data and is based on CNNs. No prior research in the field of medical big data analytics has, as far as they are aware, addressed both forms of data simultaneously. With a convergence time of only 94.8% and a prediction accuracy that surpasses that of most conventional methods, the proposed method significantly exceeds a CNN-based unimodal disease risk prediction algorithm [15].

In this study, Sahoo, Mohapatra and Wu (2016) a probabilistic data-gathering technique is created, and then the acquired data was analysed for correlation. Lastly, a stochastic prediction model is made to forecast the future health state of the most related folks based on their existing status. Extensive cloud-based simulations allow for the performance assessment of the suggested protocols; these simulations achieve a prediction accuracy of around 98% while reducing analysis time by 90% while maintaining 90% CPU and bandwidth utilisation [16].

In this study, Abdel-Zaher and Eldeib (2016) by combining the unsupervised route of a deep belief network with the supervised path of back propagation, a CAD technique for breast cancer diagnosis has been created. The architecture is a Backpropagation Neural Network (BPN-NN) trained using the Liebenberg-Marquardt learning function, with weights initially set using the Deep belief network (DBN-NN) route. They validated our method using the WBCD or Wisconsin Breast Cancer Dataset. A 99.68% accuracy rate from the classifier complex is encouraging when compared to other published research. As a breast cancer categorization model, the suggested approach works well. A number of train-test partitions were also considered when analysing the design [17].

In this study, Kandaswamy et al. (2016) are very interested in the use of state-of-the-art ML techniques, such as DNNs, to categorise substances involved in chemical MOAs. To classify compounds, image-based profiling techniques have been used, sometimes in conjunction with feature reduction techniques like factor or PCA. This article demonstrates how to classify MOAs based on cell input properties independently of treatment profiles and feature reduction techniques. Our best understanding is that this is the first use of DNN using single-cell data in this field. Additionally, they employ DTL to reduce the computationally strenuous and time-consuming process of scouring the vast parameter space of a DNN. The outcomes indicate that this method results in a 30% increase in efficiency and a 2% increase in accuracy[18].

This investigation is designed to assess an efficacy of ML models in a detection of breast cancer by employing a Breast Cancer Wisconsin (Diagnostic) dataset. A following steps of research design are shown in [Figure 1] flowchart. Data preprocessing is conducted to ensure the dataset is clean and ready for analysis, including removing duplicate entries and handling missing values. Feature scaling is implemented through standardisation to normalise the data, guaranteeing that all features contribute equitably to the model’s efficacy. The preprocessed data was then split into training (80%) and testing (20%) sets. A variety of classification models, including FFNN, RF, and DT, were ultimately used. In order to determine the most successful model for breast cancer diagnosis, key measures such as F1-score, recall, accuracy, and precision were used to evaluate each prototype.

Flowchart for Breast cancer Diagnosis The following steps of a flowchart are briefly explained below:

This study makes use of the Wisconsin Breast Cancer (Diagnostic) Data Set, which is a dataset specifically designed for this purpose. There are 569 samples in the collection, and each sample has 32 visually assessed atomic characteristics calculated from an image of a breast mass’s FNA. The distribution of benign (B) and malignant (M) tumours, as diagnosed, is displayed in [Figure 2].

A class distribution analysis in [Figure 2] shows an imbalanced dataset, with the majority class (“B”) having significantly more instances than the minority class (“M”). This imbalance can lead to model bias, causing poor performance in a minority class. In these cases, standard accuracy may not be a reliable statistic. Thus, it’s better to use evaluation metrics like precision, recall, F1-score, and AUC, which provide a more realistic evaluation of the model’s performance on imbalanced datasets.

[Figure 3s] heatmap shows the relationship between several attributes, probably taken from a dataset. A color intensity varies from dark purple to light pink, indicating the strength of correlation ranging from about 0.45 to 0.90. This kind of visualisation is useful for understanding how different variables in a dataset are related to each other, which is crucial for feature selection in model-building processes.

Data preparation in the context of breast cancer detection utilising the Wisconsin dataset entails cleaning, organising, and

standardising the data so that it may be used to develop accurate and trustworthy diagnostic models [19]. A vital part of data cleaning is removing duplicates, which makes sure that there aren’t any records that are both relevant and redundant, which could affect the accuracy of analyses and models. Key pre-processing steps are listed below:

Data cleaning and preparation include removing duplicates to make sure the data is correct and dependable for modelling or analysis.

Data points lacking a value for a particular variable in a dataset are called missing values [20]. Data analysis becomes much more difficult due to missing data points, which might cause results to be biased or erroneous.

Machine learning also makes use of the standard scaler, sometimes known as standardisation, to scale features. Each feature is averaged to a zero-variance mean using this procedure [21,3]. Although this method does not restrict the data to a certain time frame or alter its distribution, it does guarantee that the majority of data points will be close to 0. This indicates that no matter how much data is scaled, outliers will remain. Equation 1 shows the definition of standard scaling. Where: xscaled = scaled sample point x = sample point x¯ = mean of the training samples σ = standard deviation of the training samples

There are two subsets of the dataset: the training set and the testing set. The model is trained using the training set, and its performance is evaluated using the test set. The Data was divided into the 80:20.

DNNs are computer models that use a layer-by-layer architecture and a large number of neurones (node) linked together by synaptic connections (weights) [22-24]. As a result, FNNs adhere to a particular architectural arrangement in which each layer’s nodes are linked to the layer below them via forward connections [25]. The limited number of neurones in a single internal hidden layer of a FNN allows it to approximate any continuous function with an activation function that is continuous and sigmoidal in nature. The connection weights provide input that a node in a FNN can process. It is possible to calculate the mathematical output yi (excitation) of a node (node i) as (2): Where: is a total incoming connection, is an input, is a weight, is a bias, and (·) determines a range of possible values for the i-th node’s output amplitude, which is controlled by the activation function.

F1 score, recall, accuracy, and precision are important performance indicators for evaluating a model’s efficacy and helping to comprehend its predictive skills as well as finding places for development [26]. The equations of metrics, as shown in, are based on the fundamental measuring parameters of the confusion matrix, The following factors should be taken into account when measuring the parameters:

True Positive (TP): Correctly identify the presence of disease. True Negative (TN): Correctly forecast the absence of disease. False Positive (FP): Incorrectly forecast the disease is present when it’s not.

False Negative (FN): Fail to detect the disease when it is present. Accuracy: Equation 3 offers the formula for calculating the percentage of true outcomes (including TN and TP) relative to a total number for gauging accuracy: Precision: the proportion of states that were considered interesting (loaded in this case) and truly exist in that state for the purpose of measuring precision. In Equation 4: Recall: the percentage of intriguing states accurately identified as such [25]. Another name for it is Recall or Sensitivity [27]. The formula for measuring recall is mentioned in Equation 5: F1-score: a measure of both accuracy and memory that yields the proportion of correctly identified occurrences [28]. The formula for measuring F1-score are mentioned in Equation 6: The Findings are obtained by evaluating the model’s performance using these performance metrics on the testing set.

Here, the outcomes for the various classification systems used in this study are examined. Our research employed ML techniques for an effective detection of breast cancer, specifically focusing on models such as FNN that compare with RF [29] and DT [30] shown in [Table 3]. An effectiveness of these algorithms was evaluated employing the Breast Cancer Wisconsin (Diagnostic) dataset. Important performance measures utilised to assess a model’s utility were F1- score, recall, accuracy, and precision. [Table 2]. shows the results of the proposed model.

[Table 2] and [Figure 4] presents a performance of the FNN model for breast cancer prediction. The model achieved an accuracy of 96.49%, with both precision and recall scores reaching 97.18%. Additionally, the F1-score was also 97.18%, indicating a well-balanced performance according to both sensitivity and specificity. These results suggest that the FNN model is highly effective for predicting breast cancer, demonstrating strong classification performance across key metrics. [31-40]

In the [Figure 5] displays a confusion matrix for a FNN, presented in a 2x2 grid format. The matrix is labelled with “Actual” on a y-axis and “Predicted” on an x-axis, indicating two classes (0 and 1). The values within the matrix are: 41 TN and TP, representing correct predictions for classes 0 and 1, respectively. There are 2 FP and 2 FN, illustrating instances where the predictions did not match the actual classes. The matrix uses a colour gradient from light to dark blue to represent the range of values from low to high, accompanied by a colour bar on the right showing values from 0 to 60.

The [Figure 6] displays a ROC curve for a FNN. It plots the TPR against the FPR across a range of threshold values. The ROC curve, represented by a solid orange line, sharply rises close to a top-left corner of a graph, suggesting high model performance. The AUC is notably high at 0.96, indicating excellent discriminative ability. A dashed blue line, representing a random classifier’s performance, diagonally divides the plot, providing a baseline for comparison.

The FNN model’s Precision-Recall curve in [Figure 7] shows strong performance with an AUC of 0.98, indicating good identification of true positives. Initially, precision is high, but it gradually drops as recall increases, typical of precision-recall curves. The sharp drop in precision suggests potential class imbalance, as the model sacrifices precision to capture more true positives.

[Table 3]. above displays the outcomes of comparing a models’ performance. Among an algorithms tested, FNN performed best with an accuracy of 96.49%, surpassing both RF (94.11%) and DT (94.3%). The FNN’s remarkable recall and precision scores of 97.18% show that it is very good at reducing FP and increasing real positive detections. In comparison, the RF model demonstrated a precision of 94.97% and a recall of 94.11%, reflecting its effectiveness but slightly lower than that of the FNN. The Decision Tree model, while achieving a respectable accuracy of 94.3%, had lower precision at 91.0% and recall at 92.5%, suggesting it may face challenges in accurately identifying all positive cases. The FNN is the best model for detecting breast cancer overall, outperforming all other models in every metric, highlighting its potential to improve diagnostic precision in clinical settings [41-57]. Based on results its evident that AI technology may improve clinical care, education and training. However, clear regulation and understanding by clinicians are needed. ML is a subfield of AI creating systems that can improve predictions and decisions by exposure to

data, thereby imitating human learning [58,59].The integration of advanced preprocessing techniques with machine learning significantly enhances the accuracy of mammography analysis, facilitating more precise differentiation between malignant and benign breast lesions [60]. Problems such as model generalization, bias, transparency, interpretability, accountability, and data privacy remain barriers for broad adoption AI in cardiology [61].Significant potential of AI-based algorithms in enhancing the accuracy of BC survival predictions. However, further exploration and research are essential to fully understand the true impact and effectiveness of these methods [62].

Cancer is a major public health concern since it is a major killer and is on the rise around the world. Breast cancer ranks high among malignancies, particularly among females, according to current studies. Early detection can reduce treatment costs and improve survival rates for people with breast cancer. Nevertheless, the early diagnosis techniques used in modern healthcare systems have disadvantages. This study uses the Breast Cancer Wisconsin (Diagnostic) dataset to assess ML algorithms that can detect breast cancer. The results showed that Feedforward Neural Networks (FNN) were more effective than RF and DT models in detecting breast cancer. With an accuracy of 96.49% and high recall, precision, and F1-scores, the FNN proved robust in minimising false positives and maximising true positives. Nevertheless, the model’s generalisability could be compromised by obstacles including possible class imbalance and the dataset’s small size and lack of diversity. To tackle these issues, future research could use bigger and more varied datasets in conjunction with sophisticated deep-learning methods like CNNs or transformers. Additionally, integrating explainable AI methods could enhance model interpretability, facilitating its adoption in clinical settings for reliable and transparent diagnostic support.