Abstract: In the evolution of precision medicine and intelligent healthcare, integrating multimodal data including electronic health records, medical imaging and omics is critical to enhancing the reliability of clinical decision support. Nevertheless, multi-level data incompleteness has become a pervasive bottleneck restricting model performance. This paper systematically summarizes the universal statistical characteristics and modality-specific patterns of missing multimodal medical data, and analyzes its root causes from the perspectives of clinical workflows, resource constraints, hardware failures and inherent technical limitations. It examines the limitations of traditional imputation methods, and introduces three emerging fusion-based imputation paradigms: early fusion, intermediate fusion and late fusion. Finally, future research directions covering missing mechanism modeling, scalable architectures, clinical interpretability and algorithmic fairness are outlined, offering a theoretical reference for developing missingness-resilient multimodal medical learning models.