Key takeaway 📈
By lifting data, compute and visualisation into Huawei Cloud, the Servei d’Emergències Mèdiques can now anticipate call‑centre load hours or weeks ahead and re‑allocate ambulances in near real time.
The platform blends classic ARIMA with modern XGBoost/LSTM models and streams results straight into Kibana dashboards for dispatchers.
Why predict emergencies?
Between 4,000 and 70,000 incidents can hit Catalonia’s EMS on any given day, depending on ordinary seasonality—or extraordinary events like Covid waves. Manual roster planning struggled to keep up. The SEM–CIDAI project therefore set out to:
- Forecast incident volume at 1 h / 4 h / 24 h resolution up to one week ahead.
- Surface anomalies (spikes, dips) as fast as they appear.
- Provide a stand‑alone demonstrator so analysts can iterate without touching live systems.
Cloud architecture at a glance
| Layer | Huawei Cloud service | Specs / purpose |
|---|---|---|
| Compute | Elastic Cloud Server (ECS) | 8 vCPU, 32 GB RAM, 200 GB SSD + A30 GPU for training/inference |
| Public access | Elastic IP (EIP) | Dynamic BGP, up to 300 Mb/s to expose APIs & dashboards |
| Durability | Cloud Backup & Recovery (CBR) | Point‑in‑time snapshots of data & notebooks |
| Search & viz | Cloud Search Service (CSS) | 3‑node Elasticsearch (4 vCPU, 8 GB RAM each) + Kibana 7.10.2 |
The physical layout keeps data at rest on the ECS disk, GPU next to CPU, and a private VPC linking everything; only Kibana’s public URL is exposed for read‑only viewing.
Data pipeline
- Ingest & unify
Ten‑year call history—timestamp, location, incident code—plus external open data (weather, traffic, hotel occupancy, public events). - Daily ETL into CSV → Elasticsearch indices (historic + prediction).
- Model refresh on weekly, bi‑weekly or monthly cut‑offs; results are auto‑indexed so dashboards update with zero manual clicks.
Modelling strategy
“Pick the best model per horizon, not one model to rule them all.” — PT3 notes
| Horizon | Leading model | Rationale |
|---|---|---|
| 1 day | LSTM / GRU | Captures fine‑grained temporal patterns; lowest RMSE at ≤ 24 h |
| 7 days | XGBoost | More robust as error grows with horizon; handles exogenous variables well |
| Seasonal | SARIMA | Provides statistical baseline & interpretability for capacity planning |
A standard 70‑15‑15 split (train–val–test) keeps metrics honest before the chosen model is retrained on all available data and promoted to production only if it beats the incumbent.
Dashboards that dispatchers actually use
- Historic vs forecast overlay — quickly spot divergence.
- Error ribbon auto‑updates once ground truth lands to flag model drift.
- Heat‑maps by region & shift for ambulance allocation.
- Alert thresholds in Kibana Lens trigger when volume exceeds the predicted 95ᵗʰ percentile.
Because Kibana rides on CSS, analysts get full‑text search, vector similarity and Lens drag‑and‑drop without running extra servers.
Impact & next steps
Early tests show double‑digit reductions in over‑staffing during calm periods and faster escalation during spikes. The team plans to:
- Feed real‑time streaming data (not just daily batches).
- Experiment with Temporal Fusion Transformers for longer horizons.
- Offer an open API so neighbouring emergency services can plug in their data.
Final thoughts
Cloud isn’t just about elastic GPUs; it’s about turning predictions into action fast. By combining Huawei Cloud’s managed stack with a thoughtful mix of statistical and deep‑learning models, SEM’s new platform shows how public‑health services can leap from reactive to proactive—without waiting for the next crisis.
Citation (white paper): CIDAI‑PAI 2024‑02, “Predicció de Densitat d’Incidents del Sistema d’Emergències Mèdiques (SEM)”.