Modality-Specific Sparse Autoencoders for Efficient Multimodal ICU Alignment: A Symmetry–Asymmetry Learning Framework

Hashim Ali, Muhammad Tahir Akhtar

2026

Full text Open access: Gold Metadata (CSL-JSON)
DOI 10.3390/sym18040677
Data confidence 70 medium
How this score is calculated

It is built from checkable properties of the record itself: whether the article has a DOI, how many independent sources confirmed its data, whether the required fields are filled in, and how recent the latest confirmation is. Disagreement between sources and a missing identifier lower it. A value of 80 or above means the record was published without further checking; 60 to 79 means it is published at reduced confidence, with a manual check advisable. A record verified by an editor by hand is never pushed below 90 by a recalculation. This scores our data about the article, not the article itself.

Abstract

Intensive care units (ICUs) generate heterogeneous data streams, including structured electronic health records, physiological time series, and medical imaging, that describe the same patient state through different observational forms. Effective multimodal learning in this setting requires a principled balance between representation-level symmetry and architectural asymmetry. Clinically corresponding patient states should exhibit cross-modal representational symmetry, whereas each modality retains intrinsic asymmetry in dimensionality, temporal resolution, noise characteristics, and missingness. This study proposes a modality-specific sparse autoencoder framework for efficient multimodal ICU representation learning under this symmetry–asymmetry principle. Separate sparse encoders are assigned to each modality to preserve the modality-dependent structure while suppressing redundant latent activity through adaptive gating. Representation-level symmetry is encouraged through a sparsity-aware contrastive objective that aligns paired latent embeddings across modalities only on active informative dimensions. To further model inter-patient dependencies, the framework incorporates a graph neural network (GNN) whose message-passing operations respect modality-specific sparsity patterns. Experimental results indicate that the proposed framework improves predictive performance and computational efficiency relative to conventional multimodal baselines, while also exhibiting stronger robustness under missing-modality conditions and more selective latent representations. Overall, the method provides an effective and clinically relevant multimodal learning strategy for ICU decision support while offering a measurable symmetry-aware and asymmetry-preserving formulation for heterogeneous medical data.