Clinical Decision Support · CDS Hooks Ready

Identify Critical
ED Patients
Before They Crash

Sentrelia is an AI-powered triage risk stratification tool that integrates directly into Epic — delivering real-time ICU risk scores to nurses at the moment of triage, with zero additional data entry.

0.862
AUC-ROC
NHAMCS temporal holdout · 3 architectures at 0.86+
512K+
ED Encounters
NHAMCS + MIMIC-IV-ED
99.3%
Negative Predictive Value
At 95% sensitivity
0
Extra Data Entry
Reads Epic automatically
The Problem

ED Triage Misses 1 in 5 High-Risk Patients

The Emergency Severity Index (ESI) — the gold standard for ED triage across the United States — was designed for speed, not precision. Patients presenting with non-specific complaints, atypical symptoms, or early decompensation routinely receive lower acuity scores than their clinical trajectory warrants.

Under-triage leads to delayed escalation
High-risk patients assigned ESI 3–5 may wait hours before a physician assessment, during which time preventable deterioration occurs.
ESI AUC-ROC is only 0.74 for ICU prediction
Our analysis of 425,000 MIMIC-IV-ED visits confirms that standard triage acuity alone is insufficient to predict critical outcomes at the time of presentation.
The downstream cost is substantial
Avoidable ICU admissions, prolonged ED stays, and preventable deterioration represent significant financial and quality-of-care burden for health systems.

ICU / Critical Outcome Prediction

Receiver Operating Characteristic (ROC) curves — MIMIC-IV-ED anchor-era temporal holdout (PRIMARY; n = 77,026) · Outcome: ICU admission within 6 hours or ED death

Sentrelia Struct+NLP (AUC 0.943 · DeLong CI 0.939–0.946)
Sentrelia Structured (AUC 0.909)
MIMIC Acuity Alone (AUC 0.836)
Clinical Validation

Robust Performance Across Two Independent Datasets

Validated on both a nationally representative sample (NHAMCS) and a large academic medical center dataset (MIMIC-IV-ED) totaling over 512,000 emergency department encounters.

National Dataset

NHAMCS 2018–2022

National Hospital Ambulatory Medical Care Survey

0.862
AUC-ROC
Temporal holdout (2018–21 → 22)
86,864
ED Visits
146
Features
2.0%
ICU Rate
Specificity @ 80% sensitivity 88.7%
NPV @ 80% sensitivity 99.5%
Alert rate @ 80% sensitivity 13.0%
✓ Temporal holdout AUC 0.862 · Multi-architecture convergence (XGBoost 0.863, LightGBM 0.861, CatBoost 0.863)
Academic Medical Center

MIMIC-IV-ED v2.2

Beth Israel Deaconess Medical Center

0.943
AUC-ROC
Struct+NLP · 6h ICU · anchor-era
425K
ED Visits
530
Features (Struct+NLP)
4.08%
ICU Rate (6h window)
NPV @ 95% sensitivity 99.3%
Specificity @ 90% sensitivity 67.3%
ESI Baseline (AUC) 0.74 ← Sentrelia achieves 0.93+ (6h)
✓ NPV 99.3% — confidently rule out critical patients with near certainty

MIMIC-IV-ED Operating Points — Structured + NLP Model

Clinician-selectable sensitivity thresholds allow tuning the model to institutional priorities (catch more vs. flag less). Outcome: any ICU admission during hospitalization (broader definition; AUC 0.898 for any-ICU vs. 0.943 for 6-hour ICU on the anchor-era temporal holdout).

Target Sensitivity Actual Sensitivity Specificity PPV NPV Use Case
80% 80.0% 81.0% 25.5% 98.0% Busy EDs, minimize false alerts
85% 85.0% 75.3% 21.9% 98.4% Balanced approach
90% ★ 90.0% 67.3% 18.3% 98.8% Recommended default
95% 95.0% 54.2% 14.5% 99.3% Safety-first, maximum catch
Workflow

Zero Workflow Disruption

Sentrelia runs invisibly in the background and surfaces only when it matters.

01

Nurse Opens Triage Encounter

The triage nurse opens a new encounter in Epic Hyperspace as usual. No extra steps, no separate login, no new application to open.

02

Sentrelia Reads FHIR Data

Via CDS Hooks, Epic automatically sends demographics, chief complaint, and vital signs to the Sentrelia API. The model scores the patient in <200ms.

03

Risk Card Appears in Hyperspace

A color-coded CDS card surfaces inline — below or beside the standard triage documentation — with risk score, level, and the top clinical drivers. High-risk patients get an immediate alert.

Epic Integration

Truly Plug & Play
CDS Hooks + SMART on FHIR

Sentrelia uses the open CDS Hooks standard — the same protocol Epic uses for drug-interaction alerts and sepsis warnings. No custom Epic build. No HL7 interface. One URL registration and you're live.

FHIR R4 compliant — reads Patient, Observation, Encounter
CDS Hooks standard — same protocol as existing Epic alerts
SMART on FHIR app for full detailed assessment view
HIPAA-compliant audit logging for every prediction
No PHI leaves the hospital network (on-prem deploy option)
Technical Spec
GET /cds-services → Discovery POST /cds-services/ed-triage-risk → Hook
Prefetch: Patient · Observation (vitals) · Encounter
CDS Hooks Data Flow
Epic EHR
POST /cds-services/ed-triage-risk
Sentrelia API
Prefetch payload (auto-sent by Epic):
patient: Patient/{{patientId}}
vitals: Observation?category=vital-signs&_sort=-date
encounter: Encounter/{{encounterId}}
← CDS Card response (<200ms)
HIGH RISK — 42% probability
indicator: "critical" · Anticipate ICU · Alert physician
All computation on-site · No PHI transmitted externally · HIPAA-compliant
Prospective Trial

A Partnership Built for Your Health System

We are seeking a strategic health system partner for our inaugural prospective trial. A multi-site Epic environment offers the ideal opportunity to validate real-world impact at scale.

Phase 1 · Months 1–3

Silent Deployment

Model runs in observation mode alongside standard care. Predictions are logged but not shown to clinicians. Establishes baseline metrics, validates FHIR data quality, and confirms technical integration at 2–4 pilot EDs.

IRB Approval · IT Integration · Staff Training
Phase 2 · Months 4–9

Active CDS Alerts

CDS cards go live for triage nurses. Randomized stepped-wedge design across sites allows causal inference while ensuring all sites eventually benefit. Primary endpoint: door-to-ICU time for high-risk patients.

Live Alerts · Outcome Tracking · Clinician Feedback
Phase 3 · Months 10–12

Analysis & Publication

Full outcome analysis, fairness audit by demographic subgroups, and co-authored publication in a high-impact emergency medicine journal. System-wide rollout decision informed by trial results.

Co-Authorship · System Rollout · Commercial Agreement

What Your Health System Receives

First-mover advantage in AI-assisted ED triage for your health system
Co-authorship on a high-impact clinical trial publication
Reduced door-to-ICU time for high-risk patients across trial sites
Preferred pricing and data rights for commercial deployment post-trial
No upfront cost — academic partnership model for the pilot phase

Estimated Clinical Impact

Conservative projections for a large health system adoption (scaled per annual ED volume)

Additional critical patients identified annually ~4,500
Estimated reduction in door-to-ICU time ~38 min
Alert fatigue burden (% patients flagged @ 80% sens) ~19%
* Projections based on published literature on triage intervention studies. Prospective trial will measure actual impact with rigorous methodology.
Research Foundation

Built on Rigorous Science

5
Years of NHAMCS data
(2018–2022)
XGBoost
Gradient-boosted trees with
SHAP explainability
Cross-
dataset
Validated on independent
academic medical center
Fairness
Audited
Performance verified across
age, sex, and race subgroups
The Team

Clinicians Who Build Technology

Sentrelia was co-founded by emergency physicians — including a former U.S. Army medical evacuation helicopter pilot. Built from real experience triaging patients in ERs and combat zones, Sentrelia brings clinical credibility and cutting-edge AI together.

JH

J. Avery Harrell, MD

Co-Founder & CEO

Emergency medicine physician, Wilderness EMT (NOLS), former U.S. Army medevac and mountain search and rescue helicopter pilot, and Montana ski patroller. Brings firsthand experience in high-stakes emergency care across austere and clinical environments to every aspect of Sentrelia's design. Co-developed the NHAMCS and MIMIC-IV triage models underlying Sentrelia's clinical engine.

AM

Ayed Mahmoud, MD

Founder · Physician, Developer & Entrepreneur

Emergency medicine physician and developer with a passion for applying AI to solve high-stakes clinical problems and drive meaningful innovation in health systems. Co-developed the NHAMCS and MIMIC-IV triage models underlying Sentrelia's clinical engine.

AI

Sentrelia AI Engine

Clinical Decision Support Platform

XGBoost ensemble model with SHAP explainability, CDS Hooks API, FHIR R4 integration, real-time inference at <200ms, and HIPAA-compliant audit logging.

Ready to Partner?

Let's Bring This to
Your Emergency Department

We're seeking one strategic health system to partner on our inaugural prospective trial. If your system runs Epic and cares about patient safety, we'd love to talk.

Request a Technical Deep Dive

We respond within 24 hours. No salespeople, just clinicians and software developers.