SmartTriage ER

• 

  Logo

SmartTriage ER — About the Project Inspiration

Emergency rooms everywhere face the same structural problem: too many patients, too little time, and not enough structured information when doctors first meet a patient.

In many hospitals, triage still depends on rushed verbal descriptions, inconsistent forms, or brief nurse notes. This often leads to misprioritization, long waiting times, and delayed care for high-risk patients.

We were inspired by a simple question:

What if every patient could arrive at the ER with a structured anamnesis, a risk priority, and a suggested care pathway already prepared?

SmartTriage ER was born from the idea of using AI not as a replacement for doctors, but as a clinical copilot — helping healthcare teams make faster, safer, and more consistent triage decisions.

What it does

SmartTriage ER is an AI-powered emergency room triage app designed to run on a tablet or kiosk at patient intake.

When a patient arrives, they answer a short, universal questionnaire in natural language. From this input, the system automatically generates:

a structured anamnesis

a suggested medical specialty (e.g., cardiology, neurology, psychiatry, orthopedics)

a clinical priority level (Urgent, Medium, Low)

non-diagnostic hypotheses to support medical reasoning

suggested initial exams (e.g., ECG, CBC, chest X-ray)

All results are returned in a clean JSON output that can be visualized in a dashboard or integrated into hospital systems.

The system is intentionally conservative: when in doubt, it upgrades risk rather than downgrading it.

How we built it

We designed SmartTriage ER as a modular and explainable AI workflow:

Universal Intake Questionnaire A 25-question form covering:

chief complaint

symptom onset and progression

pain location and intensity

associated symptoms

medical history

medications and allergies

mental-health risk indicators

Rule-Based Clinical Brain (Explainable AI) Instead of a black-box model, we implemented transparent clinical rules inspired by real triage protocols (e.g., Manchester Triage and ESI):

specialty routing rules

priority classification rules

red-flag detection

automatic risk upgrades

LLM-Powered Reasoning Layer A large language model transforms raw patient input into:

normalized anamnesis

diagnostic hypotheses

exam suggestions

clinical justifications

Structured Output Layer All results are returned in a strict JSON schema, making the system API-ready and easy to integrate into ER dashboards.

Challenges we ran into 1) Avoiding “AI Diagnosis”

A major challenge was ensuring the system never presents itself as a diagnostic authority.

We solved this by:

framing all outputs as decision support only

using probabilistic, non-definitive language

adding explicit clinical disclaimers

2) Balancing Simplicity and Clinical Depth

We had to keep the questionnaire short enough to complete in 3–5 minutes, while still capturing enough information for meaningful triage.

This required multiple iterations to:

remove redundant questions

simplify medical language

keep only high-signal inputs

3) Risk Calibration

Determining when to label a case as Urgent versus Medium was one of the hardest problems.

We addressed this using:

red-flag symptom rules

pain-intensity thresholds

automatic risk upgrades for:

advanced age

pregnancy

chest pain

shortness of breath

Accomplishments that we're proud of

Built a complete end-to-end ER triage prototype in hackathon time

Designed a transparent, explainable triage engine instead of a black-box model

Created a universal intake questionnaire usable across medical specialties

Implemented specialty routing + priority classification + exam suggestions in one workflow

Delivered a clean JSON API output ready for real system integration

Framed AI as a clinical copilot, not a replacement for doctors

What we learned

Explainability matters in healthcare. Doctors and nurses need to understand why a system makes a recommendation.

Conservative risk handling saves lives. In medical triage, false negatives are far more dangerous than false positives.

Language is a diagnostic signal. How patients describe pain, fear, or urgency often contains as much information as formal symptoms.

AI works best as augmentation, not automation. The goal is not to replace clinicians, but to remove friction from their workflow.

What's next for SmartTriage ER

Next, we plan to:

integrate with EHR systems

add multilingual support

include vital-sign input (e.g., heart rate, oxygen saturation)

enable continuous learning from clinician feedback

add computer-vision support for trauma assessment

pilot the system in real clinical environments

Closing thought

SmartTriage ER shows how AI can be used responsibly to improve emergency care access, reduce triage errors, and support overwhelmed healthcare systems.

Smarter triage. Faster care. More lives saved. 🏥✨

Built With

• base44
• docker
• fhir
• hl7
• json
• node.js
• openai
• postgresql
• prisma
• railway
• react
• tailwindcss
• typescript
• vercel