In emergency rooms and intensive care units, clinicians rarely have the luxury of complete patient data when making life-or-death decisions. Now, researchers at Yonsei University College of Medicine have developed an artificial intelligence system that mirrors this clinical reality, delivering reliable mortality predictions even when test results are incomplete or clinical records are fragmented.

The framework, called ETF-UML (Efficient Transformer Framework for Unified Multimodal Learning), represents a fundamental departure from existing medical AI systems that typically require complete datasets or rely on statistical imputation to fill data gaps—an approach that can introduce errors and distort clinical judgment.

Performance That Improves When Data Gets Scarce

Professor Park Yurang and researcher Hong Jaesung validated their system using two major clinical databases: MIMIC-IV, containing over 190,000 records for 30-day mortality prediction and 164,000 records for 90-day mortality, and eICU, with nearly 120,000 records for 10-day in-hospital mortality prediction. ETF-UML achieved the highest AUROC (area under the receiver operating characteristic curve) across all three mortality prediction tasks when compared to existing models including GRAPE, SMIL, SHASPEC, MUSE, single-modality models, and integrated learning fusion models.

The system's advantage became even more pronounced in data-scarce scenarios that mirror real-world emergency conditions. When the observation rate of two data types dropped to just 30%, ETF-UML outperformed top competing models by up to 5.34 points on the eICU dataset. With only two data types available, the framework exceeded existing models by 4.4 to 10.7 points—precisely the conditions where reliable AI assistance is most needed but has been least available.

The framework processes three types of clinical information: structured data including patient demographics, diagnoses, and prescriptions; time-series data such as laboratory results and vital signs; and clinical text from medical records.

Computational Efficiency Opens Door to Broader Deployment

Beyond its clinical performance, ETF-UML addresses a practical barrier to AI implementation in healthcare settings: computational cost. The system requires a maximum GPU memory usage of just 0.44GB, compared to 4.70GB for MUSE, a comparable multimodal system. Total computation is reduced by more than half compared to similar models, potentially enabling deployment on standard hospital infrastructure rather than requiring specialized computing resources.

This combination of superior performance with incomplete data and dramatically reduced computational requirements could accelerate the transition of medical AI from research settings to routine clinical practice, particularly in emergency departments and ICUs where both incomplete data and resource constraints are persistent challenges.

The research was published in Information Fusion, a journal with an impact factor of 17.4, and announced on September 16, 2026. The development suggests that the next generation of clinical AI may be defined not by how much data it can process, but by how intelligently it can work with the data actually available when decisions must be made.

Sources

  1. Stable Multimodal Medical AI Developed to Function Even with Incomplete Test Results or Clinical Records — The Asia Business Daily