CareGap Trust Planner

Inspiration

A blank spot on a coverage map looks like a region with no care, but often it is just a region with no data. We kept seeing tools that turned incomplete facility records into confident recommendations, quietly hiding uncertainty inside a single score. CareGap Trust Planner was built to do the opposite — to make weak data visible and route it to human review instead of dressing it up as certainty.

What it does

CareGap Trust Planner gives healthcare planners a set of evidence-first tools:

• Identify likely care deserts from regional patterns in the facility data.
• Distinguish them from data-poor regions where the records are too thin to judge.
• Rank referral candidates by evidence-backed trust for a specific need.
• Inspect citations — the actual facility text behind each ranking.
• Save shortlists and planning scenarios, and override suspicious claims with notes.
• Prioritize records for human review when data is contradictory or missing.

How we built it

A Streamlit app deployed as a Databricks App on Free Edition, running on the provided Virtue Foundation Dataset (DAIS 2026) — 10,088 Indian healthcare facility records installed via Databricks Marketplace. We extract capability claims from fields like description, capability, procedure, equipment, specialties and source URLs, score facility-level trust, aggregate to regional planning confidence, and persist planner actions. The app ships a 2,032-facility density-weighted sample to fit the 10MB app file limit and can query the full table live via Databricks SQL. Lakebase is the target store for planner state, with a local SQLite fallback.

What makes it different

Most tools treat missing data as missing care. CareGap Trust Planner treats them as different problems. It scores how trustworthy the evidence behind each facility is, labels regions by planning confidence, and surfaces the citations and missing fields behind every recommendation. On the real data this is stark: a city can have hundreds of facilities mentioning ICU yet none with strong supporting evidence — which we surface as a likely care gap, not sufficient coverage.

Limitations

This is decision support, not medical advice. Sparse data about a region is not proof that care is unavailable. The shipped app uses a 2,032-facility sample of the provided dataset (the full table is queryable in Databricks). Every recommendation is meant for human verification before it informs a real planning decision.

Built With

• data-quality
• databricks
• databricks-apps
• databricks-free-edition
• databricks-sql
• geospatial-analysis
• healthcare
• lakebase
• postgresql
• python
• streamlit