High-performance deep learning for metastatic breast cancer classification using a minimal feature representation

Metastatic breast cancer (MBC) accounts for the majority of mortality rate, exhibiting the request for accurate and efficient predictive models to support clinical decision-making. In this study, a high-performance deep learning framework is developed for MBC classification using a minimal feature representation. We first analyze the relationships among clinical and demographic predictors to identify features that contribute most significantly to predictive performance. Based on this analysis, a novel deep forward neural network (DFNN) architecture is proposed, and the effects of hyperparameter optimization and network structure pruning are systematically investigated. The proposed framework investigates an optimized deep feedforward neural network (DFNN) for metastatic breast cancer (MBC) classification using data from the Lynn Sage Metastasis (LSM) database covering 5-, 10-, and 15-year follow-up periods. The framework systematically integrates hyperparameter optimization with a race- and ethnicity-aware feature-ranking strategy and two distinct network-pruning schemes: simultaneous linear and exponential pruning and sequential linear-to-exponential pruning. This design enables the individual and combined effects of hyperparameter optimization, feature reduction, and pruning on predictive performance to be investigated. The optimized DFNN achieves an overall classification accuracy of 0.90, demonstrating strong discriminative capability. Notably, the pruning framework maintains comparable ROC performance while reducing the input representation from 31 predictors to only four selected features. These findings demonstrate that MBC classification can retain high predictive performance with a substantially reduced feature representation, providing a compact and computationally efficient modeling framework while explicitly examining the influence of demographic-associated predictors on feature reduction. The proposed framework therefore contributes a systematic approach for developing feature-efficient DFNN models for metastatic breast cancer prediction rather than relying solely on conventional performance-oriented model optimization.