Answers are grounded in this paper and cite their sources.
Identifying the specific subtypes of a disease is a critical step in personalizing medical treatment, yet combining the massive amounts of data from different biological layers remains a major challenge. While deep learning is great at distilling biological data into latent features, it remains to be seen if these simplified features are actually enough to represent the full complexity of a disease. In this study, we compared four distinct strategies using the BRCA and ROSMAP datasets: a View-Specific Stacked Generalization Ensemble, an End-to-End Multi-Input Deep Neural Network (Late Fusion), a Two-Stage Early Fusion Network, and our proposed Hybrid Latent-Original Feature Fusion Network. We evaluated these models based on their classification accuracy, stability, and ability to handle imbalanced data. We identified the Hybrid Latent-Original Feature Fusion Network as the most effective approach. While other methods often struggled with smaller patient groups, our hybrid model achieved the highest test accuracy 87.07% for BRCA and 82.08% for ROSMAP and weighted F1-scores (0.88 and 0.82, respectively). The strength of this architecture is dual-pathway design, which balances the stability of simplified features with the original data. This ensures that vital biological clues are not lost during the compression process. Rather than replacing one with the other, our findings confirm that we achieve the best results by combining simplified features and original biological data. This establishes a new approach for building effective deep learning models in biomedical research.
No takes yet. Share an insight, caveat, or question.
Khan et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: