The research introduces Latent-Attention Masked Autoencoders (LAMAE), a model designed for learning representations from multiple cardiac data modalities. The model operates on a study-view-entity hierarchy, exchanging information directly in the latent space. Pretraining was conducted on over 1.2 million MIMIC-IV hospital stays. LAMAE demonstrates improved performance compared to modality-specific pretraining and contrastive learning baselines across tasks including in-hospital mortality, ICD-10 and DRG coding, and length of stay. The model’s performance is maintained even when only a single modality is available at test time. This indicates that modeling both intra- and inter-modal structure yields more robust representations.
Evaluation results show LAMAE outperforms existing methods across several clinical tasks. Specifically, the model achieved superior results compared to modality-specific pretraining and strong contrastive and vision-language baselines. The model was evaluated on tasks such as predicting in-hospital mortality, assigning ICD-10 and DRG codes, and estimating length of stay. The model’s ability to perform effectively with limited data (single modality) is a key strength.
The architecture utilizes a shared latent-attention module operating over the study-view-entity hierarchy. This allows the model to aggregate variable observations and handle missing modalities gracefully. The model’s training leveraged a large dataset of MIMIC-IV hospital stays, providing a robust foundation for learning complex cardiac representations.
The research highlights the importance of multimodal representation learning in medical applications. The LAMAE model demonstrates that integrating information from multiple sources can lead to improved performance and more robust representations, particularly when dealing with incomplete data.
Source: https://arxiv.org/abs/2609.12035
