Breast cancer is the most commonly diagnosed cancer worldwide, accounting for approximately 31% of all cancers in women. According to 2020 global statistics, an estimated 2.3 million new cases were diagnosed and 685,000 deaths occurred -- meaning 1 in every 8 cancers that year was a breast cancer case.
Despite rising incidence, the causes and biological mechanisms of breast cancer remain incompletely understood. Contributing risk factors include the intrauterine environment, adolescent exposures, delayed childbearing, and declining birth rates accompanying economic transitions. These cumulative factors make it harder to rely solely on lifestyle interventions.
Current clinical screening tools -- including mammography, ultrasonography, and MRI -- require specialized equipment and trained readers. Breast cancer must also be distinguished from benign conditions such as breast adenoma and cystic hyperplasia, making misdiagnosis a real concern, particularly when imaging quality or equipment is suboptimal.
These limitations have spurred interest in automated machine learning classification systems that could improve diagnostic accuracy, reduce clinician workload, shorten patient wait times, and provide a second opinion layer to complement standard imaging-based pathways.
The study used the publicly available Wisconsin Breast Cancer Dataset, which contains 569 patient records across 32 columns -- 30 physiological features derived from digitized cell nucleus images, one ID column, and one diagnosis column (benign or malignant). Of the 569 samples, 357 were benign and 212 were malignant.
Data preprocessing included filling missing values, encoding the diagnosis label as 0 (benign) or 1 (malignant), and testing each feature's distribution for normality. Since the data did not conform to a normal distribution, Spearman correlation analysis (a non-parametric method) was used instead of Pearson correlation to measure feature-target relationships.
Features with Spearman correlation coefficients above 0.7 were retained, yielding eight high-value features: perimeter_worst, radius_worst, area_worst, concave points_worst, concave points_mean, perimeter_mean, area_mean, and concavity_mean. The Wilcoxon rank-sum test then confirmed that these features showed significantly different distributions between benign and malignant tumors (p less than 0.01).
Finally, data were normalized using z-score standardization, outliers were identified using the interquartile range (IQR) method and replaced with median values, and the dataset was split 80/20 into training and test sets before model training began.
Seven classification algorithms were trained and compared. The decision tree creates a flowchart of logical rules that partition data by feature values -- simple to interpret but prone to overfitting with shallow depth settings. The random forest combines dozens of decision trees using a bagging approach, with each tree trained on a random feature subset to improve generalization; the optimal number of trees was found to be 40.
The stochastic gradient descent (SGD) algorithm minimizes the loss function iteratively using randomly selected training samples, enabling faster convergence on larger datasets. K-nearest neighbors (K-NN) classifies each new point by majority vote among its K closest neighbors in the feature space -- effective with irregular boundaries but sensitive to scale and dimensionality.
The support vector machine (SVM) finds the hyperplane that maximally separates classes in a high-dimensional space, excelling on small samples and nonlinear problems. Logistic regression estimates the probability that a sample belongs to the malignant class and was ultimately selected as the best single model, prompting its use as the base classifier for AdaBoost.
The AdaBoost-Logistic model combines logistic regression with boosting: it trains sequential logistic classifiers where each round upweights the samples misclassified in the previous round, progressively focusing the ensemble on the hardest cases. A grid search was used to tune hyperparameters and minimize training and test errors below 1%.
Across all seven models, the AdaBoost-Logistic algorithm achieved the highest performance on the test set: accuracy of 99.12%, precision of 0.99 for benign and 1.00 for malignant, recall of 1.00 for benign and 0.98 for malignant, and an F1 score of 0.99 for both classes. Its test error rate was just 0.00877 -- under 1%.
The decision tree was the weakest performer, with 92.1% accuracy and notably lower recall (0.90) for malignant tumors. This highlights a known weakness of shallow decision trees: they tend to underperform on minority classes when there is class imbalance, as was the case here with fewer malignant samples.
Mid-tier models performed in a tight cluster: random forest, SVM, SGD, and K-NN all achieved accuracies between 95.6% and 97.4%, with the logistic regression model at 98.2% emerging as the single-algorithm leader. The ROC curve area under the curve (AUC) for AdaBoost-Logistic, SGD, and SVM were all 0.9896, while the decision tree lagged at 0.9009.
The superiority of AdaBoost-Logistic is particularly evident in its very low test error (0.00877) relative to its training error (0.06153), which indicates the model generalizes well rather than memorizing training data -- a key sign of good model quality.
The authors compared their AdaBoost-Logistic model against eight published methods, all using variants of the Wisconsin dataset. Approaches ranged from standard SVM and KNN techniques (2016-2019) achieving 73%-95.6% accuracy, to a 2021 logistic regression with AUC method at 96%, and a 2022 polynomial SVM at 99.03%.
The current model's 99.12% accuracy marginally surpasses the 2022 polynomial SVM benchmark -- a meaningful step forward given that improvements at the upper end of the accuracy scale are increasingly difficult to achieve. The result confirms that combining ensemble boosting with a well-calibrated base classifier can extract additional performance beyond what any single algorithm provides.
The authors emphasize that high accuracy alone does not guarantee clinical utility. They note that AI interpretability is essential in medicine: clinicians need to understand not just what a model predicts, but why. Opaque black-box predictions can undermine informed consent and make errors harder to catch, which is why the statistical feature analysis and data visualization steps were included as interpretability components.
The feature subset selected through Spearman correlation -- dominated by geometric measurements of cell nuclei (perimeter, area, concavity) -- aligns with established pathology knowledge. Malignant tumors tend to have larger, more irregular nuclei, so the model's reliance on these features is biologically grounded and interpretable to clinicians.
The study identifies several practical problems in current breast cancer diagnosis that machine learning could address: long patient waiting times, missed diagnoses, misclassification due to image quality issues, and high workloads for laboratory pathologists. An automated pre-screening layer could flag high-risk cases for priority review.
Beyond direct diagnostic use, the feature importance analysis yields biological insights. The fact that geometric measures of cell nuclei -- particularly worst-case perimeter, radius, area, and concavity -- carry such high discriminative power reinforces what pathologists have long known: nuclear morphology is a key hallmark of malignancy.
The authors also suggest this work provides a template for applying the same approach to other cancers. The combination of normality testing, non-parametric correlation-based feature selection, outlier handling, and ensemble boosting is a generalizable pipeline, not a breast-cancer-specific one.
Limitations include the use of a single, well-curated benchmark dataset. Real-world performance may differ when applied to data collected across different hospital systems with varying imaging equipment, staining protocols, and patient populations. Prospective clinical validation would be required before deployment.
This study demonstrated that a carefully designed machine learning pipeline -- combining statistical feature selection, normalization, outlier handling, and AdaBoost ensemble boosting over logistic regression -- can classify breast tumors as benign or malignant with 99.12% accuracy on the Wisconsin dataset.
The systematic comparison of seven algorithms provides useful guidance: ensemble methods consistently outperform single classifiers, and the quality of feature selection has an outsized effect on final accuracy. These lessons apply broadly across cancer classification research.
Future work should focus on validating the model on independent, multi-site datasets to confirm generalizability, and on building explainability tools that allow clinicians to interrogate predictions at the individual patient level. Integration with electronic health record systems would be a further practical milestone toward clinical deployment.