Neural network and machine learning-based analysis of clinical associated features in breast cancer pathogenesis
Lijing Yang, Wanju WangBackground
Breast cancer remains one of the most prevalent malignancies worldwide, with complex multifactorial etiology involving genetic, environmental, and lifestyle determinants. Early identification of critical associated features within confirmed patient cohorts through computational methods could significantly enhance clinical management and clinical decision-making.
Objective
This study aims to develop and validate a neural network framework for analyzing breast cancer associated clinical features, with particular focus on interpretable feature importance analysis and statistical validation of identified associations.
Methods
We employed a multi-layer perceptron (MLP) neural network alongside baseline models (Logistic Regression, Random Forest, XGBoost) trained on comprehensive clinical and lifestyle data from 460 breast cancer patients. To optimize clinical relevance and address class imbalance, target variables were binarized and SMOTE was applied. The dataset included 17 features encompassing demographic, genetic, behavioral, and environmental factors. Model interpretability was enhanced through permutation feature importance and SHAP analysis, while statistical significance was validated via ANOVA testing and multivariable logistic regression. Performance was evaluated using standard classification metrics and Stratified 5-Fold Cross-Validation techniques.
Results
Under stratified fivefold cross-validation, the MLP achieved overall test accuracy of 0.954 (95% CI: 0.937–0.972). Logistic regression had the highest mean AUC (0.861), while Random Forest achieved accuracy of 0.972. In multivariable logistic regression, alcohol use showed the strongest independent association with high-risk severity (adjusted OR = 10.751, 95% CI: 2.545–45.422; p = 0.001).
Conclusion
The presented framework successfully identifies and validates critical breast cancer associated features, providing an interpretable computational approach for risk stratification. The integration of neural networks with statistical validation and SHAP analysis offers a robust methodology for clinical risk assessment and personalized patient management strategies.