Intervention Granularity Matters: Coherent Treatment Bundles in Counterfactual Simulation with Clinical World Models
Changing a whole documented treatment bundle made the clinical world model react more than changing any single ventilator setting.
The paper tests counterfactual edits in Clin-JEPA by holding a patient’s history fixed at 1,019 invasive-ventilation onsets. It compares isolated ventilator-setting changes with a complete configuration taken from a real patient with a similar recent trajectory. The bundle-level edit shifted the predicted next state further across all five settings, even after controlling for input-change size. The authors argue that single-component edits can describe treatment states absent from the data and may understate model sensitivity. ArXiv · AI/CL/LG's note
The paper tests counterfactual edits in Clin-JEPA by holding a patient’s history fixed at 1,019 invasive-ventilation onsets. It compares isolated ventilator-setting changes with a complete configuration taken from a real patient with a similar recent trajectory. The bundle-level edit shifted the predicted next state further across all five settings, even after controlling for input-change size. The authors argue that single-component edits can describe treatment states absent from the data and may understate model sensitivity. ArXiv · AI/CL/LG's note
score 4