🚑 Emergency Routes

Real-Time Route Optimization for Emergency Medical Services

An integrated AI-based decision-support system for optimizing emergency medical services in the province of Valencia, developed in collaboration with the Generalitat Valenciana.

---

Table of Contents

• Abstract
• Motivation
• System Architecture
• Modules
  • Route Optimization
  • Audio Triage
• Data
• Results
• Deployment
• Setup & Reproduction
• Limitations
• Team & Contributions
• References

---

Abstract

Conventional navigation systems do not account for real-time road closures, extreme weather conditions, or large-scale urban events. In emergency situations, this forces ambulances to recalculate routes reactively, losing critical minutes.

Emergency Routes addresses this through two complementary modules:

1. Route Optimization Engine — Incorporates road restrictions as exclusion polygons in OpenRouteService, with 7 predefined scenarios (Fallas, Marathon, DANA, PATRICOVA, among others) and support for multi-casualty incidents under the START/SHORT doctrine and the MANV dispatch model.

2. Automated Triage System — Transcribes emergency calls using Whisper and classifies their severity with a fine-tuned mBERT model (bert-base-multilingual-cased), reaching a macro F1 of 85.4% and perfect recall in the Severe category.

Motivation

On October 29, 2024, the Valencian Community was hit by one of the most devastating floods in recent European history (DANA). The catastrophe caused massive road closures and communication blackouts. At 17:00, a peak of 2,438 simultaneous calls was recorded at 112, with 19,821 total calls and 4,770 incidents that day — overwhelming normal operational capacity.

This project was born from a simple question: how many minutes could be saved if the CICU always had the fastest route, the most critical patient identified, and the most precise estimated time of arrival?

System Architecture

┌─────────────────┐     ┌──────────────────┐     ┌──────────────────┐
│  Citizen Portal  │────▶│  Firebase Cloud   │◀────│ Dispatch Portal  │
│  (Netlify)       │     │  (Firestore DB)   │     │ (Netlify)        │
└────────┬────────┘     └──────────────────┘     └────────┬─────────┘
         │                                                 │
         ▼                                                 ▼
┌─────────────────┐                              ┌──────────────────┐
│  Hugging Face    │                              │ OpenRouteService │
│  Spaces (AI)     │                              │ (Routing Engine) │
│  - Whisper       │                              │ - avoid_polygons │
│  - mBERT triage  │                              │ - Dijkstra / A*  │
└─────────────────┘                              └──────────────────┘

• Frontend: Static HTML portals hosted on Netlify (citizen + dispatch).
• Real-time Communication: Firebase Firestore with onSnapshot listeners.
• AI Pipeline: Whisper (small) + fine-tuned mBERT on Hugging Face Spaces (Docker).
• Routing: OpenRouteService API with dynamic closure polygons.

Modules

1. Route Optimization with Closures

The routing module determines, in real time and under road closure scenarios, which ambulance should be dispatched and which route to follow to minimize Total Operational Latency (ambulance → incident + incident → hospital).

Key features:

• 7 preloaded restriction scenarios (Fallas, Marathon, DANA, PATRICOVA, Semana Marinera, San Vicente, Ofrenda)
• Custom CSV upload for ad-hoc closures (accidents, demonstrations, roadworks)
• Filters ambulances by operational availability (shift + status)
• Haversine pre-filter (top 8 closest) → ORS travel-time ranking
• Hospital selection filtered by incident severity
• Multi-Casualty Incident (MCI) mode: START/SHORT triage + MANV dispatch doctrine
  • Automatic SAMU/SVB role assignment
  • 70% fleet cap to preserve provincial coverage
  • Hospital dispersion when ≥ 5 ambulances dispatched

2. Audio Triage System

Processes emergency calls end-to-end: audio → transcription → severity classification.

Component	Model	Details
Speech-to-Text	Whisper small (244M params)	Supports Spanish + Valencian, no GPU required
Severity Classification	mBERT (bert-base-multilingual-cased)	Fine-tuned on 504 synthetic calls, 3 classes
Medical Context	Wikipedia API	Automatic clinical term lookup, no API key needed

Why mBERT over alternatives?

• vs. BETO: native Valencian support without translation
• vs. XLM-RoBERTa: equivalent performance (macro F1 0.853) at lower computational cost
• WordPiece tokenization handles uncommon medical terminology well

Data

Source	Origin	Status
Hospital directory	Generalitat Valenciana	🔒 Confidential
Ambulance inventory (SAMU/SVB)	Generalitat Valenciana	🔒 Confidential
Road restriction layers (7 scenarios)	Valencia City Council / GVA / OSM	📦 Download from Releases
Emergency call dataset (504 calls)	Synthetic	✅ Included
Traffic intensity 2023–2025	Generalitat Valenciana	✅ Included
River flows and reservoirs	Júcar Hydrographic Confederation	✅ Included
Weather data (AEMET)	AEMET	⏳ Future development

Note: Hospital and ambulance data are confidential and provided exclusively for academic use. See Data/README.md for details on restriction layer construction.

Results

Triage Model Performance (test set, n=76)

Metric	Value
Accuracy	85.5%
Macro Precision	87.4%
Macro Recall	85.3%
Macro F1	85.4%

Per-class highlights:

• Moderate: Perfect recall (1.00) — no moderate emergency goes undetected.
• Mild: Perfect precision (1.00) — when predicted mild, always correct. Conservative errors (mild → severe) are clinically preferable.
• Severe: Recall of 0.92, with 2 errors classified as moderate (never as mild). No severe case is ever classified as mild — the most dangerous error is fully prevented.

Route Optimization (Fallas scenario demo)

Metric	No closures	With Fallas (781 closures)
Ambulance → Incident	6.9 min	6.8 min
Incident → Hospital	3.9 min	6.5 min
Total Operational Latency	11.0 min	13.0 min

The system proactively routes around all 781 closure polygons. Without it, an ambulance would discover blocked streets only upon arrival, losing critical minutes.

Deployment

The system is fully cloud-based and accessible from any device without local installation:

Component	Service	URL
Citizen Portal	Netlify	Hosted static HTML
Dispatch Portal	Netlify	Hosted static HTML
Real-time DB	Firebase Firestore	onSnapshot push updates
AI Pipeline	Hugging Face Spaces	Docker container with Whisper + mBERT

Setup & Reproduction

Prerequisites

• Python 3.9+
• ffmpeg installed on your system
• API keys for OpenRouteService and Firebase (see .env.example)

Installation

# Install dependencies (for the triage notebook)
pip install openai-whisper transformers torch requests

# Install ffmpeg (Ubuntu/Debian)
sudo apt install ffmpeg

Running the Triage Model

Open Code/audio_clasificacion_emergencias.ipynb in Jupyter or Google Colab (GPU recommended for faster transcription).

Running the Dispatch Portal Locally

Open Code/mockup/emergencias_global.html in a browser. You will need to enter your ORS API key in the interface. Firebase must be configured with your own project credentials.

Project Structure

emergency-routes/
├── README.md
├── .gitignore
├── .env.example
│
├── docs/
│   └── G1_PROYIII_LaTex.pdf          # Full project report
│
├── Code/
│   ├── audio_clasificacion_emergencias.ipynb   # Whisper + mBERT triage pipeline
│   ├── embalses_caudales.ipynb                 # Reservoir/river flow analysis
│   ├── datos_hidrograficos_chj.xlsx            # CHJ hydrographic data
│   ├── llamadas_emergencia.xlsx                # Synthetic emergency call dataset
│   └── mockup/
│       ├── ciudadano_global.html               # Citizen portal (alert submission)
│       └── emergencias_global.html             # Dispatch portal (route optimization)
│
└── Data/
    ├── README.md                               # Data documentation
    └── intensidad_total_23-25.csv              # Traffic intensity (future development)

⚠️ Road restriction scenario CSVs (7 files, ~53 MB) are not included directly in the repository due to GitHub's file size limits. Download them from the Releases section of this repo.

Limitations

• ORS API load limits: The public API cannot process the full DANA exclusion cartography at once. A self-hosted ORS instance would be needed for regional-scale disasters.
• Synthetic training data: The triage model was trained on 504 synthetic calls because real 112 recordings are confidential. Performance on real calls may differ.
• Hugging Face free tier: The AI container sleeps after inactivity; first request after sleep has additional latency.
• Flood prediction model deferred: Only one DANA event exists in the historical record, making it impossible to build a generalizable flood prediction model.

SDG Alignment

• 🏥 SDG 3 — Good Health and Well-Being: reducing emergency response latency.
• 🏙️ SDG 11 — Sustainable Cities and Communities: urban resilience to disasters.
• 🌍 SDG 13 — Climate Action: preparing emergency systems for extreme weather.

Team & Contributions

Name	Contributions
Jorge Acín Zurita	Data collection, audio triage pipeline (Whisper transcription + mBERT classification), model fine-tuning, hyperparameter search, Wikipedia medical context module, SDG alignment analysis
Pablo Arnau Tomás	Data collection and cleaning (hospitals, ambulances, traffic intensity, AEMET, CHJ hydrographic data), exploratory analysis, synthetic emergency call dataset generation, project report and documentation (LaTeX), results evaluation and visualization
Nacho Castilla Llorca	Cloud deployment (Netlify, Firebase Firestore, Hugging Face Spaces), real-time communication between portals, citizen and dispatch portal frontend development, route optimization engine, ORS integration, dispatch algorithm (Haversine pre-filter, ambulance selection)
Pablo Huguet Chordi	Data collection, project report and documentation (LaTeX), results evaluation and visualization, system testing and scenario simulations, SDG alignment analysis
David Llácer Córdoba	Data collection, geospatial data processing (QGIS), construction of road restriction layers (Fallas, Marathon, DANA, PATRICOVA, Semana Marinera, San Vicente, Ofrenda), coordinate reprojection (UTM ↔ WGS84)
Óscar Muñoz García	Data collection, audio triage pipeline (Whisper transcription + mBERT classification), model fine-tuning, hyperparameter search, Wikipedia medical context module, SDG alignment analysis, synthetic emergency call dataset generation

Note: All team members contributed to the overall project design, discussion of results, and iterative improvement across modules.

Institution: Polytechnic University of Valencia (UPV) — Bachelor's Degree in Data Science, Project III

Collaborators: Yulia Karpova and M.ª Fulgencia Villa

License

This project was developed for academic purposes as part of the Data Science degree at UPV. The source code is released for educational and research use. Confidential data provided by the Generalitat Valenciana is not included and must not be redistributed.

References

See the full reference list in docs/FinalReport.pdf, Section 8.