Paper review · October 22, 2026
TRACE: Tracking multimodal clinical trajectories to predict inpatient risk
TRACE: Tracking Multimodal Clinical Trajectories for Inpatient Risk Prediction
TRACE organizes EHR data, chest X-rays, radiology reports and discharge summaries from a hospital stay into a day-by-day trajectory and encodes it with time-ordered attention, achieving the highest AUROC across all three datasets and all three prediction tasks.
Problem
A patient's condition changes over time during a hospital stay, and heterogeneous data such as lab tests, medications, imaging and notes accumulate irregularly. Prior multimodal risk prediction research has focused on fusing modalities or compensating for missing modalities. How the evidence changes over the course of a hospital stay has received relatively less attention.
Method
Static context such as demographics and diagnosis information becomes global tokens, while each day's EHR events, chest X-rays, radiology reports and discharge summaries become local tokens for that date. Each input is encoded with a frozen medical vision or clinical language encoder, and modality and date embeddings are added. Modalities with no records produce no tokens, so no separate imputation is needed. The TGL attention mask lets every token always attend to the global tokens, while local tokens can only see the same day or earlier dates. Contrastive learning, which defines positive pairs using both outcome labels and diagnosis-based phenotypes, is added to the prediction loss.
Results
- The model predicted in-hospital mortality, prolonged length of stay and 30-day readmission on about 180,000 hospital and ICU records from MIMIC-IV, eICU and MIMIC-III. It achieved the highest AUROC in all 9 dataset-task combinations and the highest ACC in 8.
- The baselines are fusion-focused, graph- and relation-based, and missing-modality methods (MultimodalPred, AdaCoMed, MM-STGNN, M3Care, MUSE and others).
- In MIMIC-IV ablations, the TGL mask achieved higher AUROC than other mask designs on all three tasks.
- The authors report that phenotype-based contrastive learning raised the average per-phenotype AUROC in all 9 combinations.
Limitations
Evaluation used only public retrospective data (MIMIC, eICU), with no prospective clinical validation. Within the portion of the paper reviewed, specific AUROC values and the size of the gaps over baselines were not visible. Grouping data by day loses fine-grained timing within a day, and the encoders are used frozen.
Why it matters
Clinical data are inherently incomplete and irregular, so a design that uses only what was recorded, in time order, may fit practice better than forcing missing modalities to be filled in. Interpretability features that show which day and which modality changed the risk are useful for clinical decision support.
Key terms
- EHR (electronic health record)
- Structured patient records stored in hospital information systems, such as lab values, prescriptions, procedures and diagnosis codes.
- Temporal Global–Local attention
- An attention mask that always allows reference to static patient information, while restricting date-specific events to see only the same day or the past.
- Supervised contrastive learning
- A training method that pulls together the representations of samples with the same label (here, outcome + phenotype) and pushes apart those of other samples.
- AUROC
- A metric that expresses, on a scale from 0.5 to 1 and independent of the threshold, how well a binary classifier separates positives from negatives.
Original
Authors: Minjun Kim, Jong Hak Moon · Posted 2026-10-19 · License CC BY-NC-SA 4.0 · Based on the full text
This review is an AI-written summary of the paper. Please treat the original paper as the reference.