Native Key-Padding Masking and Gated Cross-Attention for Missing-Modality Multi-Omic Prognostication in Fatal Neurodegeneration

Multi-omic biomarker profiling holds transformative potential for survival modeling in fatal neurodegenerative disorders such as Amyotrophic Lateral Sclerosis (ALS). However, translational utility is systematically compromised by real-world missing-data regimes, where invasive assays or patient dropouts produce incomplete modality profiles. Conventional post-hoc imputation methods introduce substantial variance and distortion. We introduce a native architectural solution integrating Key-Padding Masking and Gated Cross-Attention directly into a multi-modal deep survival transformer. By enforcing dynamic attention redistribution across intact clinical, transcriptomic, and epigenetic channels, our model completely circumvents imputation. Evaluated across the 1,000-patient multimodal cohort from the Answer ALS consortium, our framework demonstrates robust prognostic performance under severe dropout regimes (up to 75% missing modalities), achieving superior concordance indices (C-index = 0.841) and Brier scores compared to baseline imputation and late-fusion architectures. This paradigm enables reliable prognostic risk stratification across heterogeneous clinical cohorts.