An Adaptive Framework for Breast Cancer Prognosis Using LARF-Driven Feature Selection, Self-Tuning Hyperparameters and Multi-Model Ensemble Fusion

Abstract

The use of advanced pre-processing methods, feature selection, hyperparameter optimization, and ensemble learning methods in this study provides a holistic solution for breast cancer diagnosis. The technique employs the synthetic minority over-sampling technique (SMOTE) to rebalance the dataset, along with the interquartile range (IQR) method to identify and eliminate outliers, thereby maintaining data integrity, which helps address class imbalance and makes the data more robust and representative. The feature selection strategy involves a non-linear combination of LASSO and RFE, called LASSO-RFE, to identify the most discriminative predictors. Finally, a random forest (RF) classifier is optimized using random search in order to obtain 97.23% accuracy. Finally, ensemble learning is utilized with adaptive boosting (AdaBoost), categorical boosting (CatBoost), and extreme gradient boosting (XGBoost) to achieve improved performance. The stacking ensemble approach yields strong performance with an accuracy of 97%, precision of 95%, recall of 95%, F1-score of 95%, and area under the curve (AUC) score of 96%. Such outcomes demonstrate the strong complementarity among the algorithms, while mitigating their individual weaknesses, thereby significantly improving the overall model when multiple algorithms are used. Advanced machine learning (ML) techniques therefore have huge potential to generate highly accurate and reliable predictions.