Mr.Dec: Daily-Scale Longitudinal Multimodal Modeling for 30-Day Readmission Prediction
Mr.Dec introduces a Transformer Decoder-based architecture that models hospital admissions as chronological sequences of daily multimodal events, preserving granular day-level clinical signals often lost in fixed-representation approaches The model integrates daily Electronic Health Record (EHR) updates with intermittent Chest X-ray (CXR) findings in a time-aligned stream, reflecting actual clinical workflows Disease-Specific Supervised Contrastive Learning is employed as auxiliary regularizatio
Analysis
TL;DR
- Mr.Dec introduces a Transformer Decoder-based architecture that models hospital admissions as chronological sequences of daily multimodal events, preserving granular day-level clinical signals often lost in fixed-representation approaches
- The model integrates daily Electronic Health Record (EHR) updates with intermittent Chest X-ray (CXR) findings in a time-aligned stream, reflecting actual clinical workflows
- Disease-Specific Supervised Contrastive Learning is employed as auxiliary regularization to induce diagnosis-aware structure in the latent space, improving robustness
- Evaluations on MIMIC-IV and MIMIC-CXR datasets demonstrate state-of-the-art performance in 30-day readmission prediction while preserving clinical sequence integrity
- The model identifies "Critical Days" within admissions, offering actionable and clinically interpretable insights for real-time risk stratification
Why It Matters
This work addresses a critical gap in clinical AI by moving beyond static, compressed representations of patient history to capture the dynamic, day-level evolution of clinical risk during hospitalization. For AI practitioners working in healthcare, it demonstrates how Transformer Decoder architectures can be effectively adapted for longitudinal multimodal medical data, and how contrastive learning regularization can improve generalization in clinical prediction tasks.
Technical Details
- Architecture: Mr.Dec (Multimodal Readmission-risk prediction Decoder) leverages a Transformer Decoder to process each admission as a natural chronological sequence of daily multimodal events, rather than compressing longitudinal history into fixed-length representations
- Multimodal Integration: The model time-aligns daily EHR updates with intermittent CXR imaging findings, creating a unified stream that mirrors real clinical documentation workflows
- Regularization Strategy: Disease-Specific Supervised Contrastive Learning serves as an auxiliary objective, encouraging the latent space to develop diagnosis-aware clustering that improves generalization across patient populations
- Datasets & Benchmarks: Evaluated on MIMIC-IV (EHR data) and MIMIC-CXR (chest X-ray dataset), achieving state-of-the-art results on 30-day readmission prediction tasks
- Interpretability: The model identifies "Critical Days" — specific admission days that most significantly influence readmission risk predictions — providing clinically grounded explanations for risk stratification decisions
Industry Insight
- The shift from fixed-representation to sequence-preserving architectures for longitudinal clinical data represents a promising direction for healthcare AI, particularly as real-time risk monitoring becomes increasingly valued in hospital settings
- The use of contrastive learning as regularization in medical multimodal models offers a transferable strategy for improving robustness when training data is limited or imbalanced across diagnostic categories
- The "Critical Days" interpretability feature addresses a key barrier to clinical adoption — providing actionable, time-specific explanations that clinicians can act on during real-time patient management
Disclaimer: The above content is generated by AI and is for reference only.