Multimodal Deep Stacking of Clinical and Multiomics Data for Breast Cancer Prognosis Prediction

Breast cancer prognosis prediction remains challenging due to the heterogeneous nature of clinical and multiomics data and the complex interactions among diverse biological mechanisms. This study proposes a multimodal deeplearning framework integrating clinical variables, gene expression profiles, and copy number variation (CNV) data for breast cancer prognosis prediction.

Three modality-specific models were developed: a one-dimensional convolutional neural network for clinical data, a multilayer perceptron for gene expression data, and a denoising autoencoder-based neural network for CNV data. Snapshot ensemble learning was incorporated into all base models to improve prediction diversity, while bootstrap aggregation was additionally applied to the clinical model to enhance robustness. Unimodal probability outputs were transformed into engineered interaction features and integrated using a nonlinear deep stacking meta-learner. The framework was evaluated on the METABRIC breast cancer cohort using an independent test set and comprehensive performance metrics.

The multimodal framework achieved a ROC-AUC of 0.878, 84.6% accuracy, 79.4% precision, 95.6% specificity, and 72.5% average precision. It outperformed the individual unimodal models while producing satisfactory probability calibration.

Multimodal stacking effectively leveraged complementary clinical and molecular information to improve breast cancer prognosis prediction. The proposed approach provides a flexible framework for multimodal data integration in precision oncology.