PULSEGUARD

Multi-event intraoperative early-warning · research prototype

A model that watches all vital signs together and flags 9 types of intraoperative deterioration before they occur, indicating which one and why. Built on public data (VitalDB, 5,151 surgeries) and evaluated on a locked, held-out test set — every figure below is out-of-sample, with limitations stated.

Unified early-warningany deterioration, 5 min ahead — locked held-out test

0.454 AUPRC · 0.849 AUROC

Warning lead timehow early the alarm fires, median

~4.3 min

Alarm precision (calibrated)vs ~0.12 in deployed field systems

0.62–0.71 PPV

Multi-event monitor9 events — says which one is coming

0.74–0.96 AUROC

Cross-hospital, zero-shotSeoul → UC-Irvine, no retraining

0.78–0.80 AUROC

Causal: does low BP cause AKI?mostly confounding — replicated, 2 hospitals

RR 1.77 → 1.07

Stated limitation: intraoperative hypotension is near-unpredictable here — recall ~0.12, shown 7 independent ways. Reported, not omitted.

The monitor below replays real held-out cases. Use the AI toggle to compare the model against the bedside monitor.

Aastik Rajan · Johns Hopkins University
benchmark + code: github.com/AastikRajan/pulseguard

Research prototype — not a medical device, diagnosis, or treatment recommendation.

ECG II render · visualization aid

HRbpm

ART render · visualization aid

ARTmmHg

--/--

PLETH render · visualization aid

SpO₂%

ETCO₂mmHg

RR/min

BISidx

TEMP°C

VENTILATORPCV

--PIP

--PEEP

--TV ml

--MV L

CASE TIMELINE · SpO₂ vs AI RISK — lead-time reveal

SpO₂ AI risk AI alarm Event onset

00:00 / 00:00