Reliability-Coupled Manifold-Aware Diffusion for Missing-Modality Inference

In real-world multimodal time-series inference, a common practice is completion–inference decoupling: missing modalities are heuristically imputed and then passed to deterministic fusion and decision making. We show that this chained design yields unreliable predictions under missingness, noise, and temporally varying modality reliability, because imputation errors propagate without uncertainty feedback. Through theoretical and empirical analysis, we identify latent-consistent completion and evidential reliability modeling as two key ingredients for robust multimodal inference. Building on these insights, we propose MD2E-MI, a manifold-aware diffusion-based imputation and evidence-driven multimodal inference framework that couples diffusion-based completion with uncertainty-aware inference in a single pathway. Experiments demonstrate stable performance across diverse missingness and noise regimes, while providing interpretable, temporally varying reliability estimates.