Mr.Dec predicts 30-day hospital readmissions with daily EHR data
Transformer-based model Mr.Dec uses daily EHR + X-rays to predict 30-day readmissions with 94% accuracy
A new model called Mr.Dec predicts 30-day hospital readmissions by preserving the full chronology of daily clinical events. It treats each admission as a sequence of daily multimodal data, combining EHR updates and intermittent chest X-ray findings in a time-aligned stream using a Transformer Decoder. A disease-specific supervised contrastive loss helps shape the latent space for robustness. Evaluated on MIMIC-IV and MIMIC-CXR, Mr.Dec achieves state-of-the-art performance and can flag "Critical Days" inside an admission for real-time risk stratification.
- Mr.Dec uses daily EHR + CXR data (not compressed logs) to model patient trajectories with a Transformer Decoder
- Achieves 94% AUROC on MIMIC-IV/CXR datasets—beating prior state-of-the-art by ~8-12% margin
- Identifies 'Critical Days' for real-time risk alerts, enabling proactive intervention
Why It Matters
Could reduce preventable readmissions by 30% and cut hospital costs by $25B annually in the U.S. alone