Abstract: With the explosive growth of multi-source medical data such as bio-multi-omics, medical imaging, and electronic health records, a single modality is unable to characterize the biological heterogeneity of complex diseases. Multimodal medical data fusion technology provides new possibilities for disease prediction and treatment by integrating heterogeneous information at the feature level, representation level, and decision level. This study systematically reviews the progress of fusion methodologies based on deep learning and statistical modeling in recent years, including end-to-end frameworks driven by Transformer and graph neural networks, explicit probabilistic inference supported by Bayesian and latent factor models, and new theoretical perspectives such as information bottlenecks and commonality-specificity decomposition to enhance representation effectiveness. In view of cross-modal heterogeneity and high-dimensional sparsity, this paper summarizes three types of fusion strategies, namely early, mid-, and late-stage, as well as training paradigms such as collaborative training and multi-view alignment, and discusses the role of attention mechanisms in capturing complementary information. Further combined with application cases such as cancer prognosis, biomarker discovery, drug response prediction, and clinical decision support, this paper explains the advantages and challenges of fusion models in improving prediction performance, enhancing interpretability, and fitting clinical workflows. This paper proposes future research directions for clinical implementation: building a secure and compliant federal data lake, developing a causal explainable fusion framework, and strengthening deep coupling with medical care processes to achieve a closed-loop transformation from multimodal data to precision diagnosis and treatment.

Key words: Multimodal fusion, Deep learning, Information bottleneck, Explainability, Precision diagnosis and treatment