TRACE: Tracking Multimodal Clinical Trajectories for Inpatient Risk Prediction

TRACE. Tracking Multimodal Clinical Trajectories for Inpatient Risk Prediction. By Minjun Kim and one other, from Yeji X. Published on the nineteenth of October. No code repository is linked to it.

Predicting risk for a hospital inpatient means modeling clinical evidence that is heterogeneous and that accumulates irregularly over time. The challenge covers in-hospital mortality, prolonged length of stay, and readmission, and each of these comes with its own observation window. Existing multimodal approaches already combine structured health record data, chest x-rays and clinical text. Their focus, though, has mostly been on fusing modalities, or on compensating for modalities that are missing. Much less attention has gone to how a patient's trajectory unfolds over the course of an admission.

TRACE, short for trajectory modeling of multimodal clinical events, works at a daily scale. It organizes clinical events into day-level trajectories. Within that structure it separates two kinds of information: a global context that does not change from day to day, and local events that are specific to a given day. An admission becomes a sequence of days, each carrying whatever happened on it, set against a background that holds for the whole stay. That framing puts time at the center, where earlier approaches put the question of how to merge the modalities.

Those daily events are drawn from EHR data, chest x-rays, radiology reports and discharge summaries. The trajectory is encoded with temporal attention, and a modality that is unavailable simply emits no tokens, so absence does not have to be filled in. A phenotype-guided contrastive objective further organizes admissions across heterogeneous disease distributions.

The evaluation uses cohorts from MIMIC-IV, eICU and MIMIC-III, nearly one hundred and eighty thousand stays in total, across mortality, length of stay and readmission prediction. TRACE achieves the highest AUROC in all nine dataset and task settings, and the highest ACC in eight of them, outperforming the baselines. Ablation studies support both of the design choices described above: the attention masks, and the contrastive learning.

There is also an interpretive payoff. Attribution at the level of the trajectory localizes which hospital days, and which modalities, are associated with risk, so a prediction can be traced back to when in the stay it came from.