SepsisWeave: Dynamic Organ Interaction Modeling with Active Evidence Acquisition for Early Sepsis Prediction

Early sepsis often emerges through subtle and evolving interactions across organ systems. Most predictors, however, collapse these dynamics into a single latent representation and return an opaque risk score, obscuring the contributions of individual organs and their dependencies. We introduce SepsisWeave, an interactive framework that formulates early sepsis prediction as a sequential evidence-acquisition task. At its core, DynaOrg independently encodes the temporal states of six organ systems and learns patient-specific, time-varying interactions among them, exposing local and interaction-adjusted deterioration risks as structured, queryable evidence. We further develop a three-stage training strategy combining interaction-format warm-up, teacher-guided strategy imitation, and student-sampled preference optimization to enable valid evidence acquisition. On MIMIC-IV, SepsisWeave consistently outperforms direct single-turn prediction from standardized summaries across multiple LLM backbones, yielding a mean relative AUROC improvement of 10.12%. Preliminary external validation on eICU-CRD further suggests improved robustness under distribution shift. These results suggest that explicitly modeling and selectively acquiring cross-organ evidence provides an effective and inspectable alternative to fixed end-to-end inference.