Beyond Imputation: A Semantic Unification Framework for Data and its Missingness in Multimodal Healthcare Analytics

Managing and analyzing multimodal data with pervasive missingness is a fundamental challenge in data engineering. This problem is particularly acute in domains like healthcare, where Electronic Health Records (EHRs) integrate diverse data types (structured tables, time-series, text) that are inherently incomplete at both the granular feature level and the coarsegrained modality level. Conventional strategies that either discard incomplete records or rely on error-prone imputation are illsuited for this reality, often introducing bias or a cascade of errors, especially when both missingness types co-exist. To address this challenge, We propose $\Sigma$-Care, a novel data representation and processing framework that fundamentally reframes data missingness not as a data quality defect to be repaired, but as an intrinsic semantic signal to be modeled. The core contribution is a semantic unification scheme that transforms heterogeneous, incomplete data, comprising present values, their provenance, explicit absence semantics and the task context, into a unified sequence of tokens. This representation sidesteps imputation and enables a task- and modality-aware Mixture-of-Experts (MoE) architecture to adaptively fuse these diverse information types. Extensive experiments demonstrated that $\Sigma$-Care outperforms state-of-the-art baselines across multiple clinical tasks. Further analyses underscore its superior robustness against data scarcity, severe multi-level missingness, and co-existing patterns of featureand modality-level missingness. $\Sigma$-Care presents a new data engineering perspective for handling complex data: by directly modeling the structure and semantics of the available information, including its absence, we can build more resilient and effective analytical systems.