CrisisLoad AI

Inspiration

The 2023 Türkiye-Syria earthquake killed over 50,000 people — not because hospitals didn't exist, but because nobody knew which hospitals were about to collapse under patient load until it was too late. Across the Mediterranean, the same pattern repeats: earthquakes in Italy, floods in Spain, conflict in Lebanon and Palestine, epidemics that don't respect borders. Health systems fail not from a single hospital overloading, but from cascading failures where ambulances route patients to already-saturated facilities while nearby hospitals with capacity sit underutilized. We built CrisisLoad AI to solve exactly this: predict the overload before it happens, and redirect resources while there's still time.

What it does

CrisisLoad AI is a system-level hospital overload early warning system that predicts overload probability 24–48 hours in advance across 15 hospitals in 5 Mediterranean countries (Lebanon, Palestine, Türkiye, Italy, Spain).

A trained ML model (Gradient Boosting, F1: 0.995) analyzes 7 real-time hospital metrics — bed occupancy, ICU usage, ER inflow, ambulance rates, supply levels, staffing, and infrastructure risk Gemini AI generates crisis narratives with actionable recommendations for each hospital When a hospital is flagged as overloaded, the system recommends alternative hospitals sorted by distance, available beds, and ICU capacity — prioritizing same-country facilities first Users can filter by country (an Italian coordinator doesn't need to see Palestine data) and simulate 5 crisis types (earthquake, flood, conflict, epidemic, heatwave) to run preparedness drills Light/dark mode for field usability — operations centers are dark, but mobile field teams need readable screens in daylight

How we built it

Backend: Python + FastAPI serving 9 REST API endpoints, backed by SQLite for zero-config deployment. We generated synthetic training data modeling 15 real Mediterranean hospitals across 6 crisis types (~6,000 records), then trained a Gradient Boosting Classifier achieving ROC-AUC of 1.0. Gemini API powers on-demand crisis intelligence.

Frontend: React + Vite with a premium glassmorphism dark theme. Recharts for interactive 72-hour metric timelines. A custom SVG Mediterranean map positions hospitals geographically with pulsing risk indicators. The alternative hospital recommendation engine calculates haversine distances and cross-references real-time capacity.

Key design choice: We deliberately avoided patient-level data. CrisisLoad AI operates entirely on system-level aggregated metrics — making it GDPR-compliant by design and deployable without hospital IT integration barriers.

Challenges we ran into

Realism in synthetic data: Crisis metrics don't follow normal distributions — an earthquake doesn't gradually increase ER load, it spikes. We had to carefully model each crisis type with distinct metric profiles (e.g., earthquakes spike infrastructure risk and ER inflow simultaneously, while epidemics slowly degrade staffing and supply) Cross-country coordination logic: Sorting alternative hospitals by "usefulness" required balancing multiple factors — same country first, then lowest risk, then closest distance. A simple sort wasn't enough Making predictions meaningful: A 73% overload probability means nothing without context. Integrating Gemini to translate raw predictions into human-readable crisis narratives ("ICU capacity will be exhausted within 12 hours; recommend activating mutual aid with Hospital X") made the system actually actionable

Accomplishments that we're proud of

ML model achieving ROC-AUC of 1.0 with Precision 0.998 and Recall 0.993 — the model reliably distinguishes overload scenarios End-to-end crisis simulation — click "Earthquake" and watch risk levels propagate across Türkiye and Italy hospitals in real-time, with automatic alternative hospital recommendations Country-filtered views — making the system practical for real coordinators who manage specific regions, not the entire Mediterranean The alternative hospital recommendation engine — this is the feature that could actually save lives by preventing ambulance misrouting during the critical first hours of a crisis

What we learned

Health system resilience is fundamentally a coordination problem, not just a prediction problem. Predicting overload is step one; the real value is in the "so what now?" — which is why alternative hospital recommendations and Gemini-powered action plans matter more than the ML accuracy score Synthetic data generation is an underappreciated skill — the quality of crisis-specific data distributions directly determined model performance System-level AI (operating on hospital metrics rather than patient data) is a powerful approach that sidesteps privacy concerns while still delivering actionable intelligence

What's next for CrisisLoad AI

Real hospital data integration via HL7 FHIR / OpenHIE feeds from partner hospitals Multi-language support (Arabic, Turkish, Italian, Spanish) for Mediterranean field teams Mobile-first progressive web app for on-the-ground coordinators Automated alert system — push notifications and SMS when a hospital crosses the amber/red threshold Inter-hospital resource sharing protocol — not just "go to Hospital X" but "Hospital X is sending 3 ambulances and 50 units of O-negative blood" Pilot deployment with MEDAIGENCY partner institutions in Lebanon and Türkiye

Built With

• css
• fastapi
• google-gemini-api
• javascript
• python
• react
• recharts
• scikit-learn
• sqlite
• vite