A public health team in Vastopolis, a metropolitan area of about two million residents, has observed a sharp increase in flu-like illnesses over the past few days, with several deaths believed to be associated with the current outbreak. The team wants to understand where the outbreak started and how the infection is being transmitted (person-to-person, airborne, waterborne, or something else) so that emergency management resources can be deployed appropriately.
The available data consists of: geolocated microblog messages posted from mobile devices over a period of several days, each with a timestamp, GPS coordinates, and text content; a map of the metropolitan area showing highways, hospitals, landmarks, and water bodies; population statistics for 13 city zones; and daily weather observations including wind speed and direction.
The analyst will need to inspect intermediate results visually and use domain knowledge to interpret patterns and make judgements at each stage.
Design an analytical workflow for this problem.
I'm mapping out a multi-stage pipeline that starts with data ingestion and text mining to identify illness-related posts, then moves into geospatial-temporal visualization to track symptom onset across the 13 zones. From there, I'll normalize by population density to account for reporting bias, apply clustering and hotspot detection to find outbreak epicenters, and correlate spatial patterns with environmental features like water bodies and wind patterns to infer transmission mechanisms. Each stage incorporates human-in-the-loop validation so domain experts can guide the analysis forward.
Analytical Workflow: Vastopolis Outbreak Investigation
Stage 1 – Data Preparation
- Clean & geocode microblog data; remove duplicates/bots; parse timestamps to consistent timezone.
- Join population statistics to the 13 zones (spatial join using zone boundaries).
- Align weather data (wind speed/direction) by day for later correlation.
- Checkpoint: Visual scatter of all geolocated posts on the map to confirm coverage and spot GPS errors (e.g., points outside the metro area).
Stage 2 – Text Filtering & Symptom Classification
- Apply keyword/NLP filtering (flu, fever, cough, sick, vomiting, etc.) to isolate illness-related posts.
- Optionally cluster symptom terms to distinguish respiratory vs. gastrointestinal language (clue for airborne vs. waterborne).
- Checkpoint: Analyst manually samples classified vs. unclassified posts to verify precision/recall of the filter.
Stage 3 – Spatiotemporal Exploration
- Build a time-animated density map (heatmap or hexbin) of illness-related posts by hour/day.
- Identify the earliest cluster(s) in time — candidate outbreak origin.
- Overlay hospitals, landmarks, water bodies, highways to see proximity of early cases.
- Checkpoint: Domain expert inspects animation to flag plausible index location(s) (e.g., near a reservoir, transit hub, or stadium).
Stage 4 – Population Normalization
- Compute illness-report rate per capita per zone per day (using population stats), not just raw counts.
- Rank zones by normalized incidence to correct for population density bias.
- Checkpoint: Compare raw vs. normalized hotspot maps — resolve discrepancies with analyst judgment.
Stage 5 – Transmission-Pathway Hypothesis Testing
| Hypothesis | Expected Pattern | Data to Check |
|---|---|---|
| Waterborne (point source) | Cases cluster around a specific water body/supply; sudden onset near source, then diffuse | Overlay cases on water infrastructure; check GI symptoms |
| Airborne (wind-driven) | Case density shifts downwind over time; direction matches wind vector | Cross-correlate daily wind direction with spatial shift of new cases |
| Person-to-person | Gradual spatial spread from index case along transit/commute routes; growth follows population density/mobility, not wind | Check spread along highways/transit corridors, contact-network-like diffusion |
- Visualize case centroid movement per day vs. wind vector per day (overlay arrows).
- Visualize spread rate vs. distance from highways/transit to test person-to-person diffusion.
- Checkpoint: Analyst determines which hypothesis best fits observed spatial-temporal signature; may combine (e.g., point-source origin + secondary person-to-person spread).
Stage 6 – Source Localization
- Trace back the earliest normalized cluster and cross-reference with map landmarks (e.g., convention center, water treatment plant, market).
- Validate using text content of earliest posts (mentions of specific location/event).
Stage 7 – Synthesis & Reporting
- Combine spatial, temporal, and textual evidence into a narrative: likely origin point, transmission mode, and current spread trajectory.
- Provide actionable map for emergency response (zones needing resources, quarantine perimeter suggestions based on wind/mobility spread model).
Tools suited for this: GIS/mapping (QGIS, Kepler.gl, ArcGIS), text mining (Python/NLTK or spaCy), time-series/animation (Tableau, D3, or Python+Plotly), all with iterative human review at each checkpoint above.
Möchtest du, dass ich einen der Schritte (z. B. Textklassifikation oder die Windkorrelation) konkreter mit Beispielcode ausarbeite?