DEMO · REAL MOVER CASE

One intraoperative hypotension event
attributed to the cause underneath the number.

This demo runs against real physiological case data from MOVER (UC Irvine, 55,483 anesthesia cases). We pick a representative case with a sustained MAP <65 mmHg event, render the case timeline, and attribute the endotype using the same physiological-signature heuristic an anesthesiologist uses at the bedside — the wedge-product version of LINCR's audit. The eventual learned endotype head, trained on the cause-attributed data this audit produces, is the M3-M6 deliverable.

01 · CASE TIMELINE

Case #0 minutes of intraoperative physiology.

Each line is the actual recorded value from the MOVER dataset. Pink shaded windows are detected IOH events (sustained MAP <65 mmHg for ≥5 minutes). Drug administrations are stem markers along the bottom.

Mean Arterial Pressure (mmHg)
Threshold line at MAP = 65 mmHg. IOH event windows shaded.
Heart rate & ETCO2
Heart rate (teal) and end-tidal CO2 (amber) during the case. The HR × ETCO2 joint signature is what the endotype attribution reasons on.
Heart rate (bpm) ETCO2 (mmHg)
Drug administrations & fluid record
Vasopressors (phenylephrine, ephedrine, norepi, vasopressin) above zero; sedatives (propofol) and crystalloid below zero.
Phenylephrine Ephedrine Norepi/Vasopressin Propofol Crystalloid
02 · ENDOTYPE ATTRIBUTION

Each event gets a cause, a rationale, and a treatment recommendation.

The audit produces this for every IOH event in every case the customer uploads. At the group level, these become protocol-change reports. At the clinician level, outlier identification. For QI committees and malpractice carriers, defensible documentation that QI activity actually happened.

Heuristic: bradycardia first (HR <50). Then hypovolemia (compensatory tachycardia + ETCO2 drop + SVV high or recent EBL). Then myocardial depression (no compensatory tachycardia + ETCO2 drop + recent anesthetic load). Default to vasodilation (most common cause, especially post-induction). Same decision tree a CAA reasons through at the bedside, made reproducible in code.

02B · CF2 COUNTERFACTUAL ZOOM

What the model says would have happened, under each intervention, from the same patient state.

For one event from the case above, we run our compartmentalized neural-ODE model (CF2 / patent §[0064]–[0075], 64K params, AUROC 0.972 on MOVER 15-min hypotension). The model integrates from the patient's state at event onset, then re-integrates from the same initial state under four different intervention schedules. Each curve below is the predicted MAP trajectory over the next 15 minutes.

CF2 was trained with a contrastive causal loss — Pearl's do-calculus written as a training objective. For every drug given in the data, the loss penalized the model whenever its predicted “with drug” trajectory wasn't directionally correct vs “without drug.” The model output below is what that training objective produces at inference on a held-out case.

CF2 predicted MAP — event 1 of N
15-minute window starting at event onset. Factual = predicted MAP if interventions continued exactly as recorded. Each colored line = same X0, different intervention.
Factual (recorded interventions) Phenylephrine bolus Ephedrine bolus Crystalloid bolus Propofol turn-down

Bolus sizing (matched 3-min window): phenyl 100 mcg/min, ephedrine 10 mg, crystalloid 500 ml/min, propofol zeroed. Magnitudes calibrated to median MOVER active-minute doses. Sign of the predicted ΔMAP is what the contrastive loss enforced; absolute magnitudes are sensitive to dosing assumptions, which is why each event card above carries the full ranking and the heuristic remains the categorical attribution layer for M0/M1.

03 · AGGREGATE ACROSS THE SAMPLE

Run the same audit across N cases. Site-level pattern emerges in one query.

This is the report a Premier Anesthesia medical director sees after a 90-day retrospective audit. The headline number is the endotype distribution — what's actually driving IOH at this facility, in this group's case mix. That's the protocol-change wedge: stop guessing, start changing the highest-leverage protocol first.