Multi-modal Stacked Ensemble Model for Breast Cancer Prognosis Prediction

Breast cancer (BC) is a global health challenge that affects millions of women worldwide and leads to significant mortality. Recent advancements in next-generation sequencing technology have enabled comprehensive diagnosis and prognosis determination using multiple data modalities. Deep learning methods have shown promise in utilizing these multimodal data sources, outperforming single-modal models. However, integrating these heterogeneous data sources poses significant challenges in clinical decision-making. This study proposes an optimized multimodal CNN for a stacked ensemble model (OMCNNSE) for breast cancer prognosis. Our novel method involves the integration of the Tug of War (TWO) algorithm to optimize the hyperparameters of a convolutional neural network (CNN), enhancing feature extraction from three distinct multimodal datasets: clinical profile data, copy number alteration (CNA), and gene expression data. Specifically, we employ the TWO algorithm to optimize separate CNN models for each dataset, identifying optimal values for the hyperparameters. We then trained the three baseline CNN models using the optimized values through 10-fold crossvalidation. Finally, we utilize an ensemble learning approach to integrate the models’ predictions and apply an SVM classifier for the final prediction. To evaluate the proposed method, we conducted experiments on the METABRIC breast cancer dataset comprising diverse patient profiles. Our results demonstrated the effectiveness of the OMCNNSE approach for predicting breast cancer prognosis. The model achieved high AUC, accuracy, sensitivity, precision, and MCC, outperforming traditional single-modal models and other state-of-the-art methods.