MADAT: Missing-aware dynamic adaptive transformer model for medical prognosis prediction with incomplete multimodal data

Multimodal medical prognosis prediction has shown great potential in improving diagnostic accuracy by integrating various data types. However, incomplete multimodality, where certain modalities are missing, poses significant challenges to model performance. Current methods, including dynamic adaptation and modality completion, have limitations in handling incomplete multimodality comprehensively. Dynamic adaptation methods fail to fully utilize modality interactions as they only process available modalities. Modality completion methods address inter-modal relationships but risk generating unreliable data, especially when key modalities are missing, since existing modalities cannot replicate unique features of absent ones. This compromises fusion quality and degrades model performance. To address these challenges, we propose the Missing-aware Dynamic Adaptive Transformer (MADAT) model, which integrates two phases: the Decoupling Generalization Completion Phase (DGCP), the Adaptive Cross-Fusion Phase (ACFP). The DGCP reconstructs missing modalities by generating inter-modal and intra-modal shared information using Progressive Transformation Recursive Gated Convolutions (PTRGC) and Wavelet Alignment Domain Generalization (WADG). The ACFP, which incorporates Cross-Agent Attention (CAA) and Generation Quality Feedback Regulation (GQFR), adaptively fuses the original and generated modality features. CAA ensures thorough integration and alignment of the features, while GQFR dynamically adjusts the model's reliance on the generated features based on their quality, preventing over-dependence on low-quality data. Experiments on three private nasopharyngeal carcinoma datasets demonstrate that MADAT outperforms existing methods, achieving superior robustness in medical multimodal prediction under conditions of incomplete multimodality.