MAFNet: A Missing-Aware Adaptive Fusion Model for Preclinical Alzheimer’s Disease Diagnosis Under Incomplete Clinical Data

Subjective Memory Concern (SMC) represents an important stage for early assessment of Alzheimer’s disease (AD), yet diagnosis remains difficult because structural magnetic resonance imaging (sMRI) alterations are subtle and may overlap with normal ageing. Multimodal approaches can complement sMRI with cognitive, genetic and demographic information, but they commonly assume complete clinical measurements and use fixed fusion strategies, limiting their ability to handle missing variables and adaptively integrate heterogeneous evidence. We propose MAFNet, a missing-aware adaptive fusion framework for classification under incomplete clinical data. MAFNet contains two complementary components. The missing-aware cross-modal conditional imputation module (MACI) exploits associations between imaging and available clinical information to reconstruct missing clinical features while preserving observed values. The diagnostic-path-driven dynamic fusion mechanism (DRFM) uses state-dependent routing to organize imaging and clinical evidence into a model-internal diagnostic path, allowing feature contributions to evolve during inference. Experiments on 719 ADNI subjects for SMC vs. NC and MCI vs. SMC yielded mean AUCs of 0.8992 and 0.9739, respectively. The ablation results further supported the effectiveness of both MACI and DRFM.