Apache County ranks first at 75.1/100 under the county model's chosen scoring assumptions. The score supports comparison; intervention benefits have not been measured.
D05 · Data Applications & Dashboards
Community Health Operations Intelligence
A public dashboard connecting Arizona county planning with census tract profiles, using CDC, Census, and CMS data, clustering, and stability testing.
Pythonscikit-learnClusteringStreamlit
The question
Where should a health system or public-health team invest limited outreach, mobile-clinic, care-coordination, or grant resources first?
Approach
- 01Public-data pipeline joining CDC PLACES, Census ACS, and CMS sources
- 02K-means and Ward clustering compared across two to six groups
- 03Resampling and feature-sensitivity checks on 1,730 eligible census tracts
- 04Live county scenarios, tract maps, model controls, and verifiable CSV downloads
Key findings
The tract study covers 1,765 Arizona census tracts. Complete values for five features support clustering for 1,730; 35 remain visible as excluded records.
K-means with two groups was the only candidate to pass the study's predefined separation, group-size, resampling, and feature-sensitivity thresholds. It produced groups of 1,228 and 502 tracts.
Its silhouette score was 0.436 and median resampling adjusted Rand index was 0.991. These describe separation and assignment consistency within this study; they do not establish predictive accuracy.
Explore the live project
From statewide patterns to individual tracts
The sidebar includes both county planning and census tract exploration. Controls select saved model results and filter the display; they do not retrain the models.
Data science methods
Testing whether community profiles hold together
The analysis joins CDC PLACES 2025 estimates with ACS 2019–2023 five-year data by tract identifier. Five inputs describe diabetes, frequent mental distress, lack of health insurance, poverty, and households without a vehicle.
Python, pandas, and scikit-learn support preparation and modeling. StandardScaler puts the features on comparable scales. K-means and Ward hierarchical clustering are compared across two through six groups.
Each candidate is checked using 20 subsamples containing 80% of eligible tracts, with scaling refitted each time. Adjusted Rand index measures how consistently tracts stay together. Further checks remove a feature, change the scaler, substitute income, or leave out a county.
Ward's two-group solution had a slightly higher silhouette score (0.448), but lower median resampling agreement (0.623). The selected K-means solution balanced separation with greater stability under the study's checks.
Try it in three steps
Compare the 18 Apache County tracts
- Open Explore census tracts. Keep K-means and two groups, clear the county selection, and select Apache. The display should contain 18 tracts.
- Choose a measure and compare two tracts below the map. Change to Ward to see how the assignments differ and read its model warning. Group numbers are labels, not priority ranks.
- Open Downloads & help to save the filtered CSV and study receipt. The receipt records the selected model, filters, and CSV hash.
Scope and evidence
What these results can support
The profiles describe similarities among areas. They do not predict individual health, establish causes, or determine funding eligibility. PLACES estimates use ACS inputs, so the sources are not fully independent. Source uncertainty is not propagated through clustering, and crude health rates can reflect age differences.
The county allocation simulator uses illustrative costs and reach assumptions that need local validation. Tract groups do not alter the county priority score.
The release passed 91 automated tests and GitHub checks for Python, Terraform, and the container. Public browser checks covered county scenarios, tract filters, model switching, and downloads. The app serves validated CSV outputs; PostgreSQL and dbt remain scaffolding, and researcher uploads and nationwide workflows remain proposed.
The earlier AWS exercise demonstrated recovery after a failed container health check. Those temporary resources were removed; the current dashboard runs on Streamlit Community Cloud.
From evidence to action
This project demonstrates public-data joins, data-quality checks, unsupervised model comparison, sensitivity analysis, and delivery through a tested interactive dashboard. Visitors can inspect the assumptions and download the filtered data with a verification receipt.