The causal hemodynamic audit for surgery.
Per-subject neural-ODE physiological estimator. Reads continuous intraoperative physiology — arterial waveform or NIBP cuff, ECG, pleth, capnography, drug record — and classifies the causal endotype driving each hypotension event. The data layer the field has been missing.
Each year ~30 million U.S. patients undergo surgery, and roughly half develop intraoperative hypotension at clinically meaningful thresholds (MAP <65 mmHg). The AKI, myocardial injury, prolonged stays, and SNF discharges that follow drive tens of billions of dollars in direct cost and a large share of anesthesia malpractice claims.
The largest randomized trial to date — IMPROVE-multi, n=8,520 — just confirmed that better blood-pressure targeting doesn't reduce any of those outcomes. Edwards's HPI, the field's FDA-cleared AI-driven IOH predictor, failed in both adoption and outcomes: predicting the number is about to drop, without classifying the cause underneath, didn't change what clinicians do at the bedside, and most clinicians don't use it.
The reason is structural. The same low BP number is driven by four physiologically distinct causes, each requiring a different clinical intervention. Clinicians already reason in those causal terms case-by-case. What they don't have is a record of what actually drove each event, a way to see patterns at the group level, or a way to systematically improve.
Treatment: vasopressor (phenylephrine, vasopressin)
Treatment: volume (crystalloid / colloid / blood)
Treatment: inotrope (epinephrine, dobutamine)
Treatment: chronotrope (atropine, ephedrine, pacing)
Sources: Sessler et al., IMPROVE-multi (2026). Walsh et al., Anesthesiology 2013. Bijker et al., Anesthesiology 2009.
The retrospective causal audit is the wedge. Anesthesia groups upload existing case data; we return per-case cause attribution, site-level pattern reports, clinician-level outlier identification, and the QI documentation hospital quality committees and malpractice carriers actually want to see.
Per-case cause attribution. Site-level pattern reports ("63% of last quarter's IOH-driven AKI was vasodilation-mediated, clustered in long cases with high MAC + propofol"). Clinician-level outlier identification. Defensible QI documentation. Three things shift downstream: site protocols change, individual practice patterns change, malpractice/regulatory exposure drops.
Pre-op case lists flagging which upcoming patients are likely to develop cause-X-driven hypotension, so the group can plan staffing, monitoring, and pharmacology before the patient is on the table. Operates in the planning window — physician-reviewable, no autonomous clinical action.
FDA-cleared (510(k)) decision support that tells the anesthesiologist, in real time, which cause is driving the current event and what the data supports doing about it. Predicate device: Edwards HPI (K181887).
CF2 is a 63,975-parameter compartmentalized neural-ODE physiological estimator. Three compartments: cardio (27 parameters), respiratory (8), pharmacokinetic (14). GRU encoder produces per-subject parameter estimates from streaming observable signals. Linear observer for waveform reconstruction. Frozen architecture, per-subject fit.
HPI remains the relevant scientific benchmark — it's the only FDA-cleared AI predictor in this space — even though it failed to change practice. CF2's benchmarks span both arterial-line and NIBP-only cases; HPI is structurally limited to the art-line subset, while CF2's learned per-subject weights transfer across both without retraining. The same model carries from inpatient hospital cases into ambulatory surgery centers — into a regime HPI cannot operate in.
| Dataset | Validation N | Prevalence | AUROC@15min | Brier (Platt-CV5) | Brier Skill Score |
|---|---|---|---|---|---|
| MOVER | 188,370 windows | 44.5% | 0.979 | 0.048 | ~0.81 |
| VitalDB | 47,717 windows | 41.1% | 0.944 | 0.070 | ~0.71 |
| MIMIC-IV | running | running | running | running | running |
| eICU | queued | queued | queued | queued | queued |
Methodology disclosure: prevalence is window-level (samples around hypotension events), not patient-level — standard in the published HPI literature and the rest of the IOH-prediction field. Patient-level intraoperative hypotension base rate is roughly 10–30% per Walsh et al. and Bijker et al. Brier Skill Score >0.7 is publication-grade calibration. Patent: USPTO Provisional App #64/011,899, "Compartmentalized Neural Ordinary Differential Equation System for Dynamical State Prediction," filed 2026-03-20, 109 clauses, sole inventor Anish Joseph.
First paid retrospective-audit pilot. Brier-Platt-CV5 paper submitted. Founders bridge to full-time on seed close.
Prospective risk stratifier ships under §520(o)(1)(E). 5 paying pilots, 2–3 annual contracts.
10 annual contracts. Real-world prospective study at 2 sites. 510(k) submission filed. First AMSO channel deal.
510(k) clearance. Real-time bedside CDS at lighthouse customers. 50+ contracts. ASC expansion via CNAP partnerships.
Practicing Certified Anesthesiologist Assistant — administers the vasopressors, fluids, and inotropes the model classifies the response to, on the patients in the populations the model is trained on. Sole inventor on the CF2 patent (109 clauses on compartmentalized neural-ODE systems for dynamical state prediction). Co-author on prior peer-reviewed work in neural prosthetics (Biomaterials, Bellamkonda lab GT/Emory, 2015) and developmental neuroscience (Lin Mei lab, GHSU). Erdős Institute alumnus; two more ML publications in progress with Algoverse. The 2013 ASEE paper that originated the LINCR name is mine — Georgia Tech BME, twelve years of continuity to the company name.