Kurate.org
Rankings
Back to Rankings

TRACE: Tracking Multimodal Clinical Trajectories for Inpatient Risk Prediction

Minjun Kim, Jong Hak Moon

Oct 19, 2026arXiv:2610.25765v1
cs.AI
Share
Scorecard· 16/16
5.0/10 impact

Solid, well-engineered multimodal trajectory model with broad evaluation, but incremental novelty, small AUROC margins, and baseline/leakage concerns limit impact.

Abstract

Accurate inpatient risk prediction requires modeling heterogeneous clinical evidence that accumulates irregularly over time. This challenge spans in-hospital mortality, prolonged length of stay, and readmission, each with task-specific observation windows. Existing multimodal approaches combine structured health record data, chest x-rays, and clinical text, but primarily focus on modality fusion or missing-modality compensation, with limited attention to temporal trajectories over an admission. We propose TRACE (TRAjectory modeling of multimodal Clinical Events), a daily-scale framework organizing clinical events into day-level trajectories, separating day-invariant global context from day-specific local events drawn from EHR data, chest x-rays, radiology reports, and discharge summaries. The trajectory is encoded using temporal attention, allowing unavailable modalities to emit no tokens. A phenotype-guided contrastive objective further organizes admissions under heterogeneous disease distributions. Across cohorts from MIMIC-IV, eICU, and MIMIC-III comprising nearly 180,000 total stays, we evaluate TRACE on mortality, length of stay, and readmission prediction. TRACE achieves the highest AUROC across all nine dataset-task settings and the highest ACC in eight, outperforming baselines. Ablation studies validate our attention masks and contrastive learning, while trajectory-level attribution localizes risk-associated hospital days and modalities.

AI Impact Assessments

(1 model)

Core Contribution

TRACE reframes admission-level inpatient risk prediction as modeling a day-indexed multimodal trajectory. Its three components are:

  • Token design. Clinical data becomes a sequence of tokens: category-level global tokens (demographics, diagnosis-derived context, APACHE) plus per-day local tokens for EHR descriptors, chest X-rays, radiology reports and discharge summaries. Missing modalities simply emit no tokens. Frozen pretrained encoders (EVA-X, BioClinical ModernBERT) feed a small trainable Transformer.
  • Temporal global–local (TGL) attention mask. Global tokens are always visible. Local tokens attend causally across days and bidirectionally within a day. A [TASK] token and a [PHENO] token sit at the end of the sequence.
  • Phenotype-guided supervised contrastive loss. Positives share both outcome label and ICD-chapter phenotype.

    • The tasks are mortality, early (48h) long length of stay, and 30-day readmission, evaluated on MIMIC-IV, eICU and MIMIC-III (~180k stays). The architectural novelty is modest. It is a sensible re-engineering of tokenized multimodal Transformers with structured attention masking, and supervised contrastive learning with a composite positive definition is a small twist. The "no-token for missing modality" idea is common in token-based multimodal models, though contrasting it with imputation or reconstruction in the clinical missing-modality literature is a reasonable framing.

    Methodological Rigor

    Strengths:

  • Three datasets, three tasks, patient-level splits, and bootstrap 95% CIs.
  • Task-specific observation windows are carefully defined, and Long LOS excludes discharge summaries to avoid leakage.
  • The ablations are informative: modality ablations, TGL vs All-Causal vs Full-Attention, and a phenotype-stratified contrastive ablation.

    • Concerns:

  • Baselines. Their implementations and tuning are unclear. Some baseline numbers look implausible or erratic, which suggests under-tuned reproductions rather than a fair comparison:

    • - MultimodalPred has 0.39 ACC on MIMIC-IV mortality.

    - Large-Ensemble ACC is identical to AdaCoMed's on some readmission entries.

    - M3Care's eICU Long LOS ACC is 0.77 against ~0.94 for others.

    - AdaCoMed's mF1 is 0.317 on MIMIC-IV readmission.

    The authors concede that baseline comparisons "should be read alongside" the ablations.

  • Margins. AUROC gains are often tiny and within overlapping CIs, for example +0.0009 on eICU mortality and +0.0015 on MIMIC-IV mortality. "Highest AUROC in all nine settings" is true but not statistically established. TRACE loses AUPRC and mF1 in many cells (CRL-MMNAR wins mortality AUPRC on MIMIC-IV, eICU and MIMIC-III). Larger gains appear on readmission and Long LOS, especially eICU readmission AUPRC (0.43 vs 0.24 for the next-best method, though M3Care's 0.2214 is lowest-tier and the baselines are weak there).
  • Contrastive ablation. Gains are small (mean ΔAUROC of +0.001 to +0.027) and are evaluated in phenotype-stratified test slices with small n (64 to 200 admissions) and no CIs. AUPRC is mixed, and t-SNE is only qualitative. Hyperparameters λ and τ were tuned per setting on validation, and the baselines may not have received equal tuning. Phenotype labels derive from diagnosis codes, which are also global input tokens, so some leakage-like redundancy is plausible. In MIMIC, ICD codes are assigned at discharge, which raises a leakage concern for the mortality and especially the 48h Long LOS tasks. The paper does not clearly address whether discharge-time diagnoses are used as global context in the early-prediction setting. This is a potentially serious validity concern.
  • Cohort sizes. The MIMIC-IV cohort (~14k admissions) is small because of the CXR-linkage requirement. Interpretability is limited to two anecdotal cases with no quantitative validation against clinicians or perturbation tests.
  • No external validation and no calibration analysis.

    • Potential Impact

    TRACE is a practical, reproducible recipe for multimodal EHR trajectory modeling. It would be a useful baseline for multimodal clinical prediction work. Day-level prefix inference and modality attribution are attractive for clinical decision support, but the evidence is retrospective, and the single-model-per-task setup is far from deployment. The conceptual message ("temporal organization is complementary to fusion and missing-modality handling") is mildly useful but not a paradigm shift. Prior temporal EHR transformers (e.g., BEHRT-style or event-sequence models) already treat records as event streams, so the claim that temporal trajectories are neglected is somewhat overstated, even if the multimodal CXR+notes+EHR combination under one mask is less common.

    Timeliness & Relevance

    Multimodal clinical foundation-style modeling with frozen encoders, irregular and missing modalities, and interpretability is a live topic. Using recent encoders (EVA-X, BioClinical ModernBERT) and recent baselines (CRL-MMNAR, EMNLP 2025; AdaCoMed, CVPR 2025) is timely. The paper does not explore LLM-based approaches.

    Strengths

  • Clean, well-motivated design with task-specific windows.
  • Broad evaluation across three cohorts and tasks.
  • Informative ablations for mask and modalities, which show the mask choice matters (Full-Attention is notably worse).
  • Honest reporting of mixed AUPRC and mF1 results and of limitations.
  • Day-level interpretability view.

    • Limitations

  • Incremental architecture, small and statistically unconvincing margins on many metrics.
  • Suspicious baseline behavior and possible unequal tuning.
  • Potential diagnosis-code and temporal leakage, and no external validation.
  • Phenotype benefit is small and inconsistent, and the interpretability is anecdotal.
  • No code release is mentioned (the authors are at a company), which lowers reproducibility, though the datasets are public and the setup is well specified.
  • Daily granularity limits acute-event modeling.

    • Overall

    This is a competent, incremental applied-ML paper in clinical multimodal prediction, with decent breadth and honest ablations but limited novelty and statistically thin gains. Its impact is likely to be modest, cited as a baseline or design reference within multimodal EHR work.

    Rating:5/ 10
    Significance 4.5Rigor 5Novelty 4.5Clarity 7

    Generated Oct 20, 2026

    Comparison History (0)

    No comparisons yet.