Researchers at ETH Zurich and the University of Basel report that a cardiac AI trained jointly on electrocardiograms, echocardiograms and chest X-rays ranked first on five of six hospital-stay prediction tasks. Their Latent-Attention Masked Autoencoder, or LAMAE, was pretrained on 1,267,017 stays from the MIMIC-IV database. The September 10 paper is an unreviewed preprint.
A hospital stay rarely contains the same set of tests for every patient. An electrocardiogram records the heart's electrical activity; echocardiography shows moving anatomy; a chest X-ray shows the heart and lungs from another angle. Most foundation models pretrain on each type separately and combine their outputs only for a later prediction. That arrangement cannot use one test to shape what the model learns from another during pretraining. A clinical model must also cope when a test is absent.
LAMAE starts with the tests available for a stay. It hides image patches, waveform segments and sometimes a complete ECG lead or imaging view. Separate encoders convert the visible pieces into numerical representations. A shared attention module then exchanges information among the test types, and separate decoders try to reconstruct the hidden input. This self-supervised task supplies its own training target, allowing the system to learn before mortality or diagnosis labels are added.
The source records included 800,035 12-lead ECGs, 227,835 chest X-ray studies and 7,243 echocardiography studies. For the hospital-level comparison, the researchers fine-tuned each model with 50,000 labelled samples across mortality, cardiovascular diagnosis, diagnosis-related group and length-of-stay tasks. LAMAE and an Independent MAE trained on the same data for 1,000 epochs; the Independent MAE omitted the shared attention module. ProbMED and MedSigLip served as two additional baselines. Each model received only the tests available for a stay at evaluation time.
LAMAE recorded an area under the receiver operating characteristic curve, or AUROC, of 91.59% for in-hospital mortality. The Independent MAE reached 90.51%, ProbMED 90.20% and MedSigLip 89.62%. AUROC measures how often a model ranks a positive case above a negative one; 50% is chance and 100% is perfect. LAMAE also led the cardiovascular diagnosis, severity, mortality-risk and length-of-stay tasks. MedSigLip retained the highest score for diagnosis-related group codes, at 59.39% against LAMAE's 58.92%.
The evidence comes from retrospective records in one database, not a prospective clinical trial. Echocardiograms and chest X-rays were also available for far fewer stays than ECGs, so test coverage influenced the comparisons with single-modality inputs. Transfer from multimodal pretraining was mixed when only one test type was used: it carried over most clearly for ECG, while chest X-ray results remained within seed variation. An external evaluation should now test hospitals with different equipment and patient populations, preserve naturally missing tests, and report calibration and clinician-assisted decisions as well as AUROC.
Sources
- Andrea Agostini and colleagues, arXiv, September 10, 2026. Abstract, authorship, submission date and licence.
- Full paper. Architecture, dataset, evaluation conditions, results and limitations.
- Lead image: Figure 2 from Agostini and colleagues, reproduced from the paper under CC BY 4.0.