TRACE: day-level multimodal clinical trajectories for inpatient risk (cross-source confirmed, HF + Kurate)

Source: HuggingFace daily papers 2026-10-21 and Kurate cs.AI #4 this week: TRACE: Tracking Multimodal Clinical Trajectories for Inpatient Risk Prediction, 2610.25765. Raw: raw/huggingface/2026-10-21-trace-tracking-multimodal-clinical-trajectories-for-inpatien.md, raw/kurate/2026-10-22-cs-ai.md.

TL;DR

Hospital risk models usually fuse EHR tables, chest X-rays and notes into one snapshot. TRACE organizes an admission into day-level trajectories: a day-invariant global context plus day-specific local events from EHR data, X-rays, radiology reports and discharge summaries. A temporal attention encoder lets a missing modality simply emit no tokens, and a phenotype-guided contrastive loss groups admissions with similar disease profiles. On ~180,000 stays across MIMIC-IV, eICU and MIMIC-III, TRACE gets the best AUROC in all nine dataset-task settings (mortality, length of stay, readmission) and best accuracy in eight; attribution localizes which hospital days and modalities drive risk.

Missing data becomes absent tokens, not imputed noise

Each hospital day contributes only the modalities it actually has.

flowchart LR
  G["Global context<br/><small>day-invariant</small>"] --> A["Temporal attention<br/><small>across days</small>"]
  D["Daily events<br/><small>EHR, X-ray, notes</small>"] --> A
  A --> C["Phenotype contrast<br/><small>groups similar cases</small>"]
  C --> R["Risk outputs<br/><small>mortality, LOS, readmit</small>"]
  classDef input fill:#d0ebff,stroke:#1971c2,color:#1b1b1b,stroke-width:2px
  classDef core fill:#e5dbff,stroke:#6741d9,color:#1b1b1b,stroke-width:2px
  classDef loop fill:#fff3bf,stroke:#f08c00,color:#1b1b1b,stroke-width:2px
  classDef exit fill:#d3f9d8,stroke:#2f9e44,color:#1b1b1b,stroke-width:2px
  class G,D input
  class A core
  class C loop
  class R exit

Blue is clinical input, purple encodes, amber is the training objective, green is the prediction.

Relation to the wiki

Outside the reader's core areas; included because it is the only paper in both the HF list and the Kurate top-20 this window. The design idea (absent modality = no tokens, rather than imputation) is the same token-economy principle as TokenGate's pruning. Kurate's top-20 this week shows all scores at the 1200 default with 0% win rates, so the tournament ranking is weak evidence this cycle.

Gaps

Retrospective benchmarks only; no prospective or cross-hospital deployment test.