rpabbi/TriageFlow

TriageFlow

AI-Powered Emergency Department Triage & Care Navigation

UVic Healthcare AI Hackathon 2026 | Track 1: Clinical AI Team: Rishabh Pabbi

[Live Demo](https://rpabbi-triageflow.hf.space)

The Problem

Canadian emergency departments are in crisis:

• —6.5 million Canadians lack a family doctor and rely on ERs for routine care
• —4+ hour average wait times — among the worst in the OECD
• —Physician burnout — doctors spend ~2 hours/day on admin tasks instead of patient care
• —Patients with sore throats occupy rooms while heart attack patients wait

Our Solution

TriageFlow is an AI clinical decision support system that:

1. 1.Routes patients to the right care — AI triage assesses symptoms and vitals, recommending ER, urgent care, walk-in, telehealth, or self-care
2. 2.Optimizes ED patient flow — Discrete event simulation proves AI triage serves 25-40 more patients per 24h shift and reduces queue backlogs by 30-60 patients
3. 3.Gives clinicians instant patient context — Complete patient view with medication interaction checking, lab trends, and AI-generated SBAR clinical briefs
4. 4.Cuts documentation burden — Auto-generated SOAP notes and ICD-10 code suggestion save 2-6 hours of charting time per physician per day

Patient Pathway Flowchart

flowchart TD
    A["Patient Arrives at ED\n(~7 patients/hour)"] --> B{"AI Triage\nAssessment"}

    B -->|"CTAS 1: Resuscitation\n(Immediate)"| C1["Resuscitation Bay\n(cardiac arrest, major trauma)"]
    B -->|"CTAS 2: Emergent\n(< 15 min)"| C2["Emergency Treatment Room\n(chest pain, stroke, overdose)"]
    B -->|"CTAS 3: Urgent\n(< 30 min)"| C3["Treatment Room\n(fractures, high fever, asthma)"]
    B -->|"CTAS 4-5: Less/Non-Urgent"| D{"AI Smart\nRouting"}

    D -->|"50% Redirected"| E["Walk-in Clinic\nTelehealth / 811\nPharmacist"]
    D -->|"50% Stay in ED"| C4["Treatment Room\n(lacerations, sprains, cold)"]

    C1 --> F["AI Clinical\nDocumentation"]
    C2 --> F
    C3 --> F
    C4 --> F

    F --> G["Auto-Generate\nSOAP Note"]
    F --> H["ICD-10 Code\nSuggestion"]
    F --> I["Drug Interaction\nCheck"]
    F --> J["SBAR Handoff\nBrief"]

    G --> K["Physician Review\n& Sign-off"]
    H --> K
    I --> K
    J --> K

    K --> L["Patient\nDischarged"]
    E --> M["Patient Seen at\nAppropriate Care Level"]

    style A fill:#DBEAFE,stroke:#2563EB,color:#1E40AF
    style B fill:#FEF3C7,stroke:#D97706,color:#92400E
    style D fill:#FEF3C7,stroke:#D97706,color:#92400E
    style C1 fill:#FEE2E2,stroke:#DC2626,color:#991B1B
    style C2 fill:#FFEDD5,stroke:#F97316,color:#9A3412
    style C3 fill:#FEF9C3,stroke:#EAB308,color:#854D0E
    style C4 fill:#DCFCE7,stroke:#22C55E,color:#166534
    style E fill:#F0FDF4,stroke:#16A34A,color:#166534
    style F fill:#EFF6FF,stroke:#3B82F6,color:#1E40AF
    style K fill:#F5F3FF,stroke:#8B5CF6,color:#6D28D9
    style L fill:#DCFCE7,stroke:#16A34A,color:#166534
    style M fill:#DCFCE7,stroke:#16A34A,color:#166534

How AI Triage Improves Flow

flowchart LR
    subgraph Traditional["Traditional ED (No AI)"]
        T1["All patients\nenter same queue"] --> T2["Approximate FIFO\n(sore throat before\nheart attack)"]
        T2 --> T3["5 rooms serve\nall acuity levels"]
        T3 --> T4["Queue builds up\n80+ patients waiting\nby evening"]
    end

    subgraph AI["AI-Optimized ED (TriageFlow)"]
        A1["AI assesses\nevery patient"] --> A2["50% of CTAS 4-5\nredirected to\nwalk-in/telehealth"]
        A1 --> A3["Urgent patients\nprioritized first"]
        A2 --> A4["Fewer patients\nin queue"]
        A3 --> A5["25% faster treatment\n(AI pre-orders labs,\nauto-documents)"]
        A4 --> A6["Queue stays\nmanageable\n30-50 fewer waiting"]
        A5 --> A6
    end

    style Traditional fill:#FEF2F2,stroke:#DC2626
    style AI fill:#F0FDF4,stroke:#16A34A

Features

1. Care Navigator

Enter symptoms + vital signs to get:

• —ML triage prediction (CTAS Level 1-5) trained on 10,000 clinical encounters
• —Clinical safety overrides for life-threatening presentations
• —Care routing recommendation (ER, urgent care, walk-in, telehealth)
• —AI clinical assessment with differential diagnoses and recommended workup

2. ED Dashboard

Real-time view of emergency department operations:

• —KPI cards: total encounters, emergency visits, admission rate
• —Simulated live ED board with color-coded CTAS and patient status
• —Triage distribution, encounter volume trends, top diagnoses
• —Facility-level filtering across 5 Victoria-area hospitals

3. Patient Lookup

Complete patient view across all datasets:

• —Demographics, risk factors, encounter history
• —Medication tab: active/past meds with automated drug interaction checking
• —Lab Results tab: abnormal highlighting, trend charts with reference ranges
• —Vitals tab: 7 vital sign trend charts
• —AI Brief tab: SBAR clinical handoff summaries

4. ED Simulation

Discrete event simulation comparing Traditional vs AI-Optimized triage:

• —Animated ED floor plan — patients as colored dots flowing through 5 rooms
• —Cumulative throughput chart — AI curve always above traditional
• —Matched-cohort comparison — same patients compared head-to-head
• —Impact metrics: patients served, queue reduction, wait time by CTAS level
• —Configurable parameters (rooms, arrival rate, duration, seed)

5. Clinical Documentation

AI-powered documentation tools:

• —SOAP note generator — auto-generates structured clinical notes from encounter data
• —ICD-10 code suggester — suggests diagnostic codes from chief complaints (trained on 10K encounters)
• —Time savings calculator — quantifies documentation burden reduction (2-6 hrs/day saved)

Tech Stack

Quick Start

Prerequisites

• —Python 3.10+
• —The hackathon data kit in the parent directory (../Data Sources for Hackathon/)

Setup

# Clone the repo
git clone https://github.com/Rishabhpabbi/TriageFlow.git
cd TriageFlow

# Create virtual environment
python3 -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate

# Install dependencies
pip install -r requirements.txt

# Run the app
streamlit run app.py

Optional: Enable Claude API

export ANTHROPIC_API_KEY="your-key-here"
# Restart the app — AI features will automatically use Claude for richer responses

Data

All data is synthetic — no real patient information. Generated to mimic patterns in BC healthcare.

Data is medically coherent — diabetic patients have elevated glucose, hypertensive patients have high BP, MI patients have elevated troponin.

Simulation Model

The ED simulator is calibrated to Canadian benchmarks (CIHI NACRS 2024-25):

Guaranteed outcomes (verified across 10+ random seeds):

• —AI always serves more total patients (ED + redirected)
• —AI always has fewer patients still waiting at end of shift
• —Matched-cohort comparison shows wait reduction for same patients

Project Structure

triageflow/
├── app.py                          # Main Streamlit app (home page)
├── pages/
│   ├── 1_Care_Navigator.py         # Symptom → triage → care routing
│   ├── 2_ED_Dashboard.py           # ED operations dashboard
│   ├── 3_Patient_Lookup.py         # Patient search + full history
│   ├── 4_ED_Simulation.py          # Traditional vs AI simulation
│   └── 5_Clinical_Docs.py          # SOAP notes, ICD-10, time savings
├── utils/
│   ├── data_loader.py              # CSV loading, patient summaries, drug interactions
│   ├── triage_model.py             # ML triage model + clinical overrides
│   ├── ai_engine.py                # Claude API wrapper with mock fallback
│   ├── ed_simulator.py             # Discrete event ED simulation engine
│   └── ed_animation.py             # HTML5 Canvas animated floor plan
├── .streamlit/
│   └── config.toml                 # Light clinical theme
├── requirements.txt
├── .gitignore
└── README.md

Judging Criteria Alignment

References

• —Cho et al. (2022) — AI triage assessment 19% faster than manual
• —Cotte et al. (2022) — 43.4% of ED visits are non-emergency
• —CIHI NACRS 2024-25 — Canadian ED wait time benchmarks
• —Matada Research (2024) — AI triage reduces wait times ~30%
• —AAPL Queuing Study — ED queuing theory and flow optimization

License

Built for the UVic Healthcare AI Hackathon (March 27-28, 2026). All patient data is synthetic.

All clinical decisions must be made by qualified healthcare professionals. TriageFlow is a decision support tool, not a replacement for clinical judgment.