MIT Laboratory for Computational Physiology Boston University School of Public Health

Watch the mechanism happen

One synthetic patient, three synchronized views. Adjust the MAP, vasopressor, and dose controls on the left and watch the causal graph light up and the state vector's cells flip between clean and treatment-confounded — both respond live to the same underlying mechanism proven on the formal proof page. The predictability gauge is separate: it shows that variant's own real, measured AUROC from Experiment 2, so switch variants below to see it move. (This is ordinary A(t-1) → S(t) treatment history shaping state, not the state containing information about the action it's used to predict — see Terminology.)

Controls

Blood pressure (MAP)68 mmHg
Vasopressor active
State variant shown to the model

1 — Causal graph (which paths are active right now)

True physiology U_t unobserved MAP_t = 68 sofa_cardio_t = 0 reflects MAP only State S_t Next action A_t Prior action A_(t-1) no vasopressor

2 — State vector (what the model actually sees)

clean physiology treatment-confounded right now

3 — Predictability (Experiment 2's measured AUROC for this variant)

—
0.50 — random guessing1.00 — perfect prediction

Formal proof →