PaperlayerView exact revision on arXiv

Source-linked AI summary

Mr.Dec: Daily-Scale Longitudinal Multimodal Modeling for 30-Day Readmission Prediction

Minjun Kim, Jong Hak Moon

arXiv:2608.16929v1cs.LG

TL;DR

Thirty-day readmission prediction needs to capture evolving, day-level clinical trajectories rather than compressing hospitalization history into fixed representations. Mr.Dec models chronological daily EHR and intermittent CXR events with a causal decoder and achieves best overall multimodal performance, while identifying clinically grounded critical days.

  • Problem

    Existing readmission models often compress longitudinal hospitalization history, limiting day-level tracking of evolving clinical risk relevant to pre-discharge screening.

  • Method

    Mr.Dec uses a causal Transformer to interleave daily EHR records and intermittent same-day CXR observations, with disease-specific supervised contrastive learning.

  • Results

    AUC 0.814 and F1 0.752 are the best overall multimodal results, surpassing MM-STGNN by +0.014 AUC and +0.184 F1.

  • Takeaways & Limitations

    Mr.Dec provides temporally grounded critical-day signals for clinically grounded real-time risk stratification.

  • Takeaways & Limitations

    MuST comparison metrics rely on the original publication because publicly available source code is absent, and F1 scores were unavailable.

Abstract

from arXiv · show

Predicting 30-day hospital readmission is essential for assessing patient stability and optimizing healthcare resources. As clinical risk evolves with the accumulation of evidence during hospitalization, capturing these dynamic trajectories is essential. However, many existing approaches compress the complex longitudinal history into fixed representations, often losing the granular, day-level clinical signals that reflect a patient's evolving physiological state. To address this, we propose Mr.Dec (Multimodal Readmission-risk prediction Decoder), which models each admission as a natural chronological sequence of daily multimodal events. By leveraging a Transformer Decoder, Mr.Dec integrates daily Electronic Health Record(EHR) updates and intermittent Chest X-ray(CXR) findings in a time-aligned stream, reflecting the actual clinical workflow. To ensure robustness, we utilize Disease-Specific Supervised Contrastive Learning as an auxiliary regularization to induce a diagnosis-aware structure in the latent space. Evaluations on the MIMIC-IV and MIMIC-CXR datasets show that Mr.Dec achieves state-of-the-art performance by preserving the integrity of the clinical sequence. Furthermore, our model identifies "Critical Days" within an admission, providing actionable and clinically grounded interpretations for real-time risk stratification. Code is available at: https://github.com/yejix-ai/MR.DEC

1 Introduction

Mr.Dec addresses the loss of day-level clinical dynamics in 30-day readmission modeling by representing hospitalizations as chronological streams of daily EHR records and intermittent same-day CXR observations. The approach emphasizes objective physiological signals and preserves intraadmission temporal structure for risk prediction.

  • Motivation: 30-day all-cause readmission prediction supports healthcare-quality assessment and early identification of high-risk patients before discharge.Unplanned readmissions worsen outcomes and impose substantial economic burden.
  • Limitations of Existing Methods: Static administrative variables such as ICD codes lack the granularity needed to represent complex clinical progression.Later approaches incorporate clinical notes and CXR through multimodal learning.
  • Multimodal Clinical Signals: Daily EHR measurements and CXR provide objective snapshots for tracking evolving physiological status at the day level.Numerical signals offer immediate quantitative indicators, while CXR can reveal internal pathology such as pulmonary edema.
  • Limitations of Existing Methods: Existing systems such as MM-STGNN and MuST often aggregate variable-length admissions into static representations, weakening intra-patient temporal granularity and causal dependencies.This limitation motivates modeling the admission as a chronological sequence.
  • Mr.Dec: Mr.Dec introduces a daily-scale causal Transformer that interleaves longitudinal EHR records with sporadic same-day CXR observations in a chronological token stream.The architecture is designed to preserve intraadmission temporal structure and jointly model the two modalities.

2 Methods

Mr.Dec formulates 30-day readmission prediction as causal modeling of a chronological, day-level multimodal admission trajectory. It interleaves daily EHR and available same-day CXR representations in a causally masked Transformer decoder and jointly trains readmission prediction with disease-specific supervised contrastive learning.

  • Trajectory formulation: Each admission is represented as a chronological sequence of hospital-day records containing structured EHR and an optional same-day chest radiograph.The model predicts an admission-level outcome from the full in-hospital history while treating unavailable CXR observations as empty.
  • Causal decoding: A causal Transformer decoder processes the admission stream with [READMIT] and [SupCon] summary tokens, whose masked attention preserves chronological information flow.The appended summary tokens can attend to the entire admission stream while respecting temporal order.
  • Multimodal tokenization: Frozen clinical and vision encoders produce EHR and CXR features, serialized as one EHR token per day with a CXR token inserted immediately afterward when imaging is available.This creates a day-aligned variable-length stream that reflects intermittent CXR acquisition.
  • Multimodal tokenization: Each token is projected into the model dimension and augmented with learnable modality and hospital-day embeddings.The resulting sequence is denoted Femb := (h1, . . . , hL).
  • Training objective: The model jointly optimizes binary-cross-entropy readmission prediction and disease-specific supervised contrastive learning using separate prediction and projection heads.Contrastive positives share the most frequent coarse ICD-10 diagnostic supergroup, among 14 total categories, to structure representations and improve robustness to noisy labels.

3 Experiments

Experiments on linked MIMIC-IV and MIMIC-CXR admissions show that Mr.Dec outperforms task-matched baselines in multimodal readmission prediction, including under class imbalance and across diagnostic categories. Its extended-window variant remains competitive on longer trajectories, while Critical Days provide temporally grounded interpretability for risk stratification.

  • Dataset: The dataset contains 13,821 adult admissions from 11,972 patients, split into 12,438 training and 1,383 test admissions without patient overlap.Admissions required a stay of at least 48 hours and at least two CXRs.
  • Overall performance: Mr.Dec achieves AUC 0.814 and F1 0.752 in the multimodal CXR–EHR setting, exceeding MM-STGNN by +0.014 AUC and +0.184 F1.The comparison uses modality masking without retraining.
  • Overall performance: Mr.Dec improves Rec(N-C) to 0.548 versus 0.329 for its BCE-only ablation while maintaining comparable AUC of 0.814 versus 0.809.The result indicates improved sensitivity to non-critical readmission cases under label noise and class imbalance.
  • Outcome-stratified performance: Mr.Dec preserves Rec(C) 0.911 and improves Rec(N-C) to 0.548, whereas LVLM baselines achieve near-zero Rec(N-C) of 0.007–0.038.Mr.Dec (Max 30-days) remains competitive on longer sequences with AUC 0.793 and F1 0.736.
  • Category-stratified performance: Mr.Dec maintains F1 0.756–0.782 across five diagnostic groups, while Mr.Dec (Max 30-days) achieves F1 0.739–0.769 across groups.The groups include radiographic-correlate and EHR-predominant conditions.
  • Interpretability: Critical Days are hospital days where token-level attributions peak and predicted risk changes most markedly, grounding real-time risk stratification in the admission timeline.Figure 2 backprojects gradient-based importance from the [READMIT] token to EHR tokens and CXR patch embeddings.

4 Conclusion · Disclosure of Interests

Mr.Dec frames readmission prediction as daily-scale causal trajectory modeling by jointly integrating longitudinal EHR and intermittent CXR data in a causal decoder. The framework combines disease-specific contrastive learning and token-level interpretability for phenotype-aware modeling and temporally grounded risk stratification.

  • 4 Conclusion: Mr.Dec redefines readmission prediction as a daily-scale causal trajectory modeling task.
  • 4 Conclusion: The framework jointly models longitudinal EHR and intermittent CXR data in a single causal decoder.
  • 4 Conclusion: Mr.Dec outperforms graph-based and large foundation model baselines, particularly for identifying critical deterioration trajectories.
  • 4 Conclusion: Disease-Specific Supervised Contrastive Learning captures inter-patient phenotypic structure without explicit graph construction.
  • 4 Conclusion: Token-level interpretability provides temporally grounded signals for real-time risk stratification.
  • 4 Conclusion: Future work will extend the multimodal token stream to additional clinical modalities such as ECG and clinical notes.
  • Disclosure of Interests: The work was supported by the TIPS Program of the Ministry of SMEs and Startups and Korea Startup Foundation under Grant No. RS-2025-25467010.
Loading 2608.16929v1…