Ambulance Vitals — Anomaly Detection Pipeline

Production-oriented pipeline for early-warning anomaly detection and triage risk scoring from streamed vitals (HR, SpO2, ABP).

Quick start

Default (recommended): run the published GHCR image

docker pull ghcr.io/keanesc/time-series-analysis:latest
docker run -p 8000:8000 ghcr.io/keanesc/time-series-analysis:latest

Verify the service:

curl http://localhost:8000/health

Manual installation (from source)

This project uses pixi for environment management. Use the commands below when developing locally or running from source.

# clone the repo (SSH or HTTPS)
cd time-series-analysis
pixi install
pixi run serve  # open http://localhost:8000
pixi run all    # run full pipeline

Core commands

• pixi run all — run entire pipeline end-to-end
• pixi run collect — fetch & simulate motion
• pixi run clean — artifact detection + cleaning
• pixi run train — train anomaly model
• pixi run score — compute risk scores
• pixi run evaluate — metrics & latency
• pixi run failures — failure-case analysis
• pixi run serve — start API server

API (essentials)

• GET /health — liveness
• POST /predict — input: hr, spo2, sbp, dbp, motion returns: anomaly_flag, risk_score, confidence, alert_status

Example:

curl -X POST http://localhost:8000/predict -H "Content-Type: application/json" -d '{"hr":120,"spo2":88,"sbp":85,"dbp":55,"motion":0.1}'

Response:

{
  "anomaly_flag": true,
  "risk_score": 78.3,
  "confidence": 0.87,
  "alert_status": "ALERT"
}

Docker

Build locally:

docker build -t time-series-analysis .
docker run -p 8000:8000 time-series-analysis

Published image (GHCR):

docker pull ghcr.io/keanesc/time-series-analysis:latest
docker run -p 8000:8000 ghcr.io/keanesc/time-series-analysis:latest

The repository includes a GitHub Actions workflow (.github/workflows/publish-ghcr.yml) that builds and publishes the image to GHCR on pushes to main. (Image is public.)

Data Source

MIMIC-III Waveform Database (PhysioNet, v1.0) [1]

• 67,830 record sets from ~30,000 ICU patients
• Numerics records at 1 Hz: HR, SpO₂, ABP (systolic/diastolic)
• Open access under ODbL v1.0
• Motion/vibration signal is simulated (ambulance context adaptation)

Model Approach

1. Clinical features (per 60 s window): distance-to-threshold, slope, mean, std, end value for each vital, plus motion and data quality indicators (24 features total)
2. HistGradientBoostingClassifier trained on stable windows to predict pre-critical deterioration within 2 minutes
3. Risk score (0–100): weighted vital severities + trend bonus, gated by data-quality confidence
4. Alert logic: ALERT if risk ≥ 70 and confidence ≥ 0.6; SUPPRESSED if confidence too low

See report.md for detailed methodology, metrics, and safety-critical analysis.

References

[1] PhysioNet, “MIMIC‑III Waveform Database (v1.0).” [Online]. Available: physionet.org/content/mimic3wdb/1.0

[2] P. Zwerschke, “Shipping conda environments to production using pixi,” QuantCo Tech Blog, Jul. 11, 2024. [Online]. Available: tech.quanto.com/blog/pixi-production