This project focuses on leveraging supervised learning models to develop an accurate breast cancer prediction system. By analyzing different features extracted from breast cancer patients, including known cases labeled as Malignant or Benign, the aim is to assess and compare various algorithms such as logistic regression, support vector machines, decision trees, random forests, SVM, and Naive Bayes. Evaluation metrics like accuracy, precision, and recall will guide the assessment process. The results of this study hold significant promise in advancing breast cancer diagnosis and improving medical decision-making, emphasizing the importance of early detection and precise diagnosis for enhanced treatment outcomes.
The dataset is obtained from kaggle and is available in this repository dataset as well.
The Kaggle breast cancer dataset consists of 569 rows and 32 columns, each representing features relevant to diagnosing breast cancer. Each entry is identified by a unique "id" and accompanied by a "diagnosis" label indicating whether the breast mass is Malignant (cancerous) or Benign (non-cancerous). Features like "radius_mean," "texture_mean," "perimeter_mean," and "area_mean" provide insights into the size, texture, and extent of the mass. Other features include measures of smoothness, compactness, concavity, and concave points, with variations for "worst" and "standard error" values. Additional features such as symmetry and fractal dimension enrich the dataset, offering insights into the characteristics of breast masses. The "diagnosis" column has two unique values, while other columns exhibit varying numbers of unique values, reflecting the diversity of measurements extracted from breast cancer data.
In the data cleaning and preparation phase, we verified that the dataset was in suitable condition for analysis. Most variables retained their original data types, with the exception of the Diagnosis target variable, which had to be transformed into binary values (1 for Malignant and 0 for Benign) to facilitate classification tasks later in the project. We had ensured that the dataset contained no null values, eliminating potential sources of error or bias in the analysis. Additionally, we had confirmed the absence of duplicate entries, ensuring the integrity and reliability of the dataset for subsequent stages of analysis. These meticulous steps laid a solid foundation for our exploration of various supervised learning algorithms and the development of an accurate breast cancer prediction model.
- The dataset exhibits significant variation in the mean radius of breast masses, ranging from 6.98 to 28.11 units, indicating diverse sizes among instances.
- The mean area of breast masses spans a wide range from 143.50 to 2501.00 square units, showcasing substantial diversity in size and extent.
- The average smoothness of breast masses varies between 0.0526 and 0.1634, suggesting a broad spectrum of surface textures from smoother to rougher.
- Breast masses also show variability in mean concavity, with values ranging from 0 to 0.4268, indicating differences in shape complexity.
- The mean fractal dimension varies between 0.04996 and 0.09744, reflecting differences in geometric patterns and intricacy. These findings underscore the importance of considering multiple features when analyzing and predicting breast cancer, as variations in size, area, texture, concavity, and fractal dimension provide valuable insights for diagnosis and further exploration.
There were 62.7% Benign (Noncancerous) cells and 37.3% Malignant (Cancerous) cells in the dataset.
The data is normally distributed for most of the columns except for radius_se, parimeter_se, area_se, and compactness_se which are right skewed.
- Radius of the Malignant tumours (Cancerous cells) is larger than the Benign (Noncancerous cells) this suggests that cancerous cells tend to grow and spread, resulting in tumors with a larger overall size. In contrast, benign cells typically exhibit a smaller radius, indicating a more confined and localized growth pattern
- Tissue average of the Malignant (Cancerous cells) is larger than the Benign (Noncancerous cells) this implies that cancerous cells tend to occupy a greater extent of the tissue, potentially indicating more aggressive and invasive behavior. Benign cells, on the other hand, occupy a relatively smaller area within the tissue.
- Perimeter thickness of the Malignant (Cancerous cells) is larger than the Benign (Noncancerous cells) this indicates that the boundary or edges of cancerous tumors are typically more irregular and spread out, while benign cells tend to have a smoother and more well-defined perimeter.
- The area occupied by Malignant (Cancerous cells) is larger than Benign (Noncancerous cells).
- Compactness_mean for Malignant (Cancerous cells) is higher which indicates that the points on the mass are more closely packed together as compared to the Benign (Noncancerous cells).
- Concavity mean for Malignant (Cancerous cells) is larger than the Benign (Noncancerous cells) which indicates a greater severity of concave regions within the mass of Malignant cells.
- Concave point mean is much higher for the Malignant (Cancerous cells) indicating that there are more concave regions as compared to Benign (Noncancerous cells).
- The outliers have been ignored as they do not make any difference in the prediction
- A significant positive correlation with a coefficient of 0.997855 was found between the variables “radius_mean” and “perimeter_mean” this is because an increase in the radius of a tumor (radius_mean) would generally result in an increase in its perimeter (perimeter_mean). This direct relationship leads to a strong positive correlation.
- The variables “radius_mean” and “area_mean” displayed a strong positive correlation, with a coefficient of 0.987357. This is because “area_mean” and “perimeter_mean” both depend on the size and shape of the tumor. As the tumor’s size increases, both its area and perimeter are likely to increase, leading to a strong positive correlation between the two variables.
- A robust positive correlation was observed between “radius_mean” and “radius_worst,” with a coefficient of 0.969539 this is because both variables capture the overall size or extent of the tumor. so, it is expected that as the mean radius of a tumor increases, the worst (largest) radius measurement would also tend to increase.
- The variables “area_mean” and “perimeter_mean” showed a strong positive correlation, with a coefficient of 0.986507. This shows that when area of tumor increases the perimeter will also increase which is quite obvious.
- A substantial positive correlation was detected between “area_mean” and “area_worst,” with 14 a coefficient of 0.959213.
- A strong positive correlation of 0.883121 was found between “compactness_mean” and “concavity_ mean”. This tells us that tumors that exhibit higher compactness tend to have more concave regions, resulting in a positive correlation between these variables. This correlation could indicate that tumors with a higher degree of compactness are more likely to have a greater level of concavity.
- A moderate negative correlation, with a coefficient of -0.311631, was observed between “radius_ mean” and “fractal_dimension_mean”. A negative correlation between these variables could indicate that tumors with larger cell radii tend to have less irregular or more uniform shapes, resulting in a lower fractal dimension.
The data problem in this study was to accurately classify breast cancer cases based on a dataset containing various tumor characteristics. The goal was to develop a predictive model that could effectively distinguish between benign and malignant tumors. To analyze breast cancer classification, this study employed several machine learning algorithms, including Random Forest, Decision Tree, Logistic Regression, KNeighbors Classifier, SVM, and Naive Bayes. These algorithms were selected due to their effectiveness in binary classification tasks. The dataset was preprocessed by label encoding the target variable, “diagnosis,” and performing feature scaling using the StandardScaler function to ensure compatibility across the algorithms. The dataset was then split into training and testing sets with a 70:30 ratio. Each algorithm was trained on the training set and evaluated on the testing set using accuracy, precision, and recall metrics.
Under the Modelling heading, the study implemented various supervised learning algorithms to classify breast cancer cases based on tumor characteristics.Each algorithm was trained and evaluated using accuracy, precision, and recall metrics. Logistic Regression and SVM exhibited the highest recall and accuracy scores, indicating their ability to correctly identify malignant cases, which is crucial in breast cancer prediction due to the high cost of false negatives. Given their comparable performance, either Logistic Regression or SVM was selected for further exploration.
Based on the evaluation results, the Logistic Regression algorithm demonstrated the highest performance in terms of accuracy and recall, making it the algorithm of choice for further analysis. The feature importance of the logistic regression model was assessed by examining the absolute coefficients, with higher magnitudes indicating stronger influences on the breast cancer prediction.
By employing logistic regression, the coefficients associated with each feature were determined, showcasing their impact on the likelihood of breast cancer presence. Positive coefficients, like those for radius_se, perimeter_worst, and concave points_worst, suggest an increase in these features correlates with a higher likelihood of breast cancer. Conversely, negative coefficients, such as texture_mean, smoothness_mean, and symmetry_se, indicate an increase in these features corresponds to a lower likelihood of breast cancer. The magnitude of the coefficients also plays a crucial role, with larger magnitudes indicating stronger influences on the prediction. Among these features, radius_se emerges as the most influential, given its highest magnitude and positive coefficient. Radius_se quantifies the standard error of the mean of distances from the center to points on the perimeter of a breast mass, thus serving as a significant predictor for breast cancer classification in logistic regression models.
- The top-left cell (106) represents the number of true negatives (TN). It indicates that there are 106 instances that were correctly predicted as the negative class.
- The top-right cell (2) represents the number of false positives (FP). It indicates that there are 2 instances that were incorrectly predicted as the positive class when they were actually negative.
- The bottom-left cell (3) represents the number of false negatives (FN). It indicates that there are 3 instances that were incorrectly predicted as the negative class when they were actually positive.
- The bottom-right cell (60) represents the number of true positives (TP). It indicates that there are 60 instances that were correctly predicted as the positive class.
The study evaluated multiple machine learning algorithms for breast cancer classification, with Logistic Regression emerging as the top performer, achieving 97% accuracy and 95.2% recall. The "radius_se" feature was identified as most influential, indicating its significant role in prediction. A confusion matrix illustrated the model's reliability with minimal false positives and negatives. Logistic Regression proved effective for early detection and diagnosis, holding promise for improved patient outcomes. The project aimed to develop an accurate breast cancer prediction model, highlighting variations in tumor characteristics and emphasizing the importance of considering multiple features. Logistic Regression's success underscores its potential as a valuable diagnostic tool, though further enhancements are possible.