Machine Learning and Feature Selection in Breast Cancer Diagnosis

Breast cancer remains a critical health concern globally, with early detection essential to improving treatment outcomes and increasing survival rates. In recent years, machine learning (ML) has emerged as a powerful tool in the medical field, enhancing diagnostic accuracy by offering refined predictive models that help identify relevant features for distinguishing between malignant and benign cases. A recent study published in Healthcare Analytics explores a comparative analysis of ML models enhanced by feature selection techniques such as Least Absolute Shrinkage and Selection Operator (LASSO) and SHapley Additive exPlanations (SHAP). Both techniques serve to optimise model performance, ultimately improving predictive accuracy, interpretability and clinical utility in early cancer detection.

Feature Selection Techniques in Breast Cancer Prediction

Feature selection is fundamental to effective machine learning models for breast cancer detection, as it helps isolate the most informative variables from extensive datasets. The study applied two feature selection methods—LASSO and SHAP—to enhance the dataset’s relevance by retaining the most impactful predictors. LASSO, a regression-based technique, effectively reduces dataset dimensionality by penalising less significant features, thus refining the dataset to include only the most crucial attributes. This reduction aids the models in concentrating on the most relevant information, reducing noise and enhancing predictive power. On the other hand, SHAP adds a layer of interpretability by assigning importance scores to individual features, allowing clinicians and researchers to better understand the weight and impact of each variable on model predictions. SHAP’s unique contribution lies in its consistent, model-agnostic approach, which quantifies feature relevance and provides explanations that are beneficial in clinical contexts where transparency and interpretability are paramount. Combining these two methods enabled this study to develop machine learning models that achieve high diagnostic accuracy, focusing on a dataset enriched by relevant, informative variables.

Comparative Analysis of Machine Learning Models

To evaluate the effectiveness of feature selection in breast cancer prediction, the study assessed several machine learning models, including logistic regression, decision tree classifiers and ensemble meta-models (such as Hard Voting and Soft Voting classifiers). These models were tested on a merged dataset combining the Wisconsin Breast Cancer Dataset (WBCD1 and WBCD2) to evaluate their performance both with and without feature selection for optimisation. Among the ML models tested, ensemble models (Hard Voting and Soft Voting classifiers) demonstrated the highest levels of accuracy and reliability, with SHAP-based feature selection propelling them to an impressive accuracy of 99.82%. Traditional models, including logistic regression and decision trees, also showed notable improvements when integrated with feature selection, underscoring the importance of isolating impactful predictors for improved results. These findings emphasise the effectiveness of SHAP in handling complex correlations between features and refining model performance, as SHAP’s ability to provide a granular view of variable influence leads to better model adjustments. The enhanced accuracy achieved across traditional and meta-models demonstrates the potential for reliable, high-performing predictive models for breast cancer detection, especially when using integrated feature selection methodologies.

Practical Applications and Future Implications

The clinical implications of integrating machine learning with sophisticated feature selection techniques extend beyond predictive accuracy. By demonstrating the effectiveness of feature selection and model tuning, this study supports the development of accessible diagnostic tools that could optimise clinical workflows. When embedded within a healthcare setting, these models offer the potential to complement traditional diagnostic methods, enabling a faster, more efficient identification process that could improve early detection rates. Early and accurate diagnosis is essential in breast cancer care, where timely intervention can significantly improve patient outcomes. With high-performing models like those developed in this study, clinicians could receive real-time support in identifying at-risk patients, ultimately enhancing decision-making and treatment planning. Future research could further expand upon this study by incorporating larger and more diverse datasets, thus addressing variations across populations and improving the generalisability of the models. Exploring deep-learning methods and other advanced machine-learning techniques could yield even higher diagnostic accuracy. At the same time, the integration of these tools into user-friendly platforms or web applications could help facilitate adoption in clinical settings, making these innovations practically accessible to healthcare providers.

The study has illustrated the importance of feature selection in advancing machine-learning models for breast cancer diagnosis. By leveraging both LASSO and SHAP methodologies, researchers were able to identify critical predictors and refine the diagnostic accuracy of traditional and meta-models alike. These advancements not only demonstrate the capacity for feature selection to enhance model performance but also highlight the broader potential of machine learning to transform cancer diagnostics. The significant improvements in diagnostic accuracy underscore the importance of combining model interpretability with predictive power, a balance that could bridge the gap between technological advancements and clinical utility. The high accuracy and interpretability achieved with these methods pave the way for future innovations in cancer diagnostics, contributing to earlier intervention, personalised treatment options and, ultimately, better consequences.

Source: Healthcare Analytics

Image Credit: iStock