Recent advances in feature selection methods for breast cancer recurrence prediction: A systematic review.

e13664 Background: Breast cancer recurrence prediction using genomic data faces challenges due to high dimensionality and complex biological relationships. This review evaluates recent methodological advances in feature selection approaches for improving prediction accuracy while preserving biological relevance. Methods: We systematically searched SCOPUS, Web of Science, and PubMed databases and identified studies published from 2018 to 2024 focusing on feature selection methods for breast cancer recurrence prediction. We analyzed four novel feature selection methods that outperform conventional univariate and multivariate approaches. These include Outlier-based Gene Selection (OGS), Semi-Supervised Survival Laplacian Regression (S3LR), pathway-based classification, and an ensemble systems biology approach. These methods were evaluated using standardized genomic datasets to compare their performance in terms of prediction accuracy, dimensionality reduction, computational efficiency, and biological significance. Results: The OGS method achieved the highest subtype-specific accuracy (F1 score: 1.0 for basal subtype) while reducing features by 91%. S3LR demonstrated superior handling of censored data (C-index: 68.80%) with 95% dimensionality reduction. The pathway-based approach maintained 85% accuracy across subtypes with 98% feature reduction. The ensemble method showed balanced performance (F1: 0.94) with 68% feature reduction. All methods demonstrated significant enrichment in cancer-related pathways (p < 0.001). Comparative analysis revealed method-specific strengths: OGS excelled in outlier detection, S3LR in handling censored data, pathway-based in biological interpretation, and ensemble in network analysis. Conclusions: Modern feature selection methods demonstrated better accuracy with fewer features compared to traditional statistical and machine learning approaches. Future studies focusing on the clinical application of these methods for real-time prediction are needed.