CareLens AI-From messy healthcare data to trusted decisions

CareLens AI by Lakehouse Lifelines transforms messy healthcare data into evidence-backed insights with confidence scores, helping planners identify real care gaps and make trusted decisions.

Comment

CareLens AI: From Messy Healthcare Data to Trusted Decisions

Team: Lakehouse Lifelines

Inspiration

Healthcare planning is only as good as the data behind it. Yet, in many regions, healthcare facility information is fragmented, inconsistent, and difficult to trust.

A facility may claim to offer ICU services, emergency care, or specialized procedures, but the supporting evidence is often buried in unstructured descriptions, incomplete records, or contradictory data fields.

This creates a critical challenge for healthcare planners, NGOs, and coordinators:

How do you distinguish between a true gap in care and a gap in data?

We built CareLens AI to answer that question.

Our goal was not to create another dashboard. We wanted to build a trust layer that helps decision-makers confidently identify healthcare access gaps while understanding the limitations of the underlying data.

What It Does

CareLens AI transforms 10,000 messy healthcare facility records into evidence-backed, uncertainty-aware insights.

The application enables users to:

  • Evaluate whether facility claims are supported by evidence
  • Identify high-risk medical deserts across regions
  • Distinguish between underserved areas and data-poor areas
  • Explore facility capabilities with source citations
  • Review confidence scores and uncertainty explanations
  • Save notes, overrides, and planning scenarios

Every recommendation includes:

  • Supporting evidence snippets
  • Source references
  • Confidence scores
  • Human review workflows

We believe users should never have to trust AI blindly.

How We Built It

CareLens AI was built entirely on the Databricks platform.

Data Engineering

  • Ingested and profiled the healthcare facility dataset
  • Standardized structured fields including location, specialties, and facility metadata
  • Cleaned noisy and repetitive text fields
  • Created a unified evidence model across descriptions, procedures, equipment, and source URLs

AI Pipeline

We treated all extracted fields as claims rather than ground truth.

Using LLM-powered extraction and classification workflows, we:

  1. Extracted facility capabilities from unstructured text
  2. Mapped evidence to standardized healthcare concepts
  3. Generated trust signals for each capability
  4. Calculated confidence scores based on evidence quality, consistency, and completeness
  5. Identified contradictory or sparse records requiring human review

User Experience

The Databricks App provides a simple workflow for non-technical users:

  1. Explore healthcare coverage on an interactive map
  2. Filter by geography and care need
  3. Review trust-weighted facility recommendations
  4. Inspect supporting evidence and uncertainty indicators
  5. Save notes and planning decisions

Databricks Technologies Used

  • Databricks Apps
  • Delta Lake
  • Unity Catalog
  • Mosaic AI
  • Vector Search
  • Databricks SQL
  • MLflow
  • Lakehouse Monitoring

Challenges We Faced

Messy Data Is Harder Than Missing Data

The dataset was not simply incomplete—it was inconsistent.

The same capability could appear across multiple fields with varying terminology, conflicting descriptions, and uneven evidence quality.

Trust Is More Important Than Accuracy

Early prototypes produced confident answers that lacked sufficient evidence.

We realized that planners need transparency more than precision.

Instead of optimizing only for prediction quality, we prioritized:

  • Explainability
  • Evidence traceability
  • Honest uncertainty

Designing for Non-Technical Users

Healthcare planners do not want to inspect model outputs or prompts.

We focused on creating a workflow that answers practical questions:

  • Where are the care gaps?
  • How certain are we?
  • What evidence supports this conclusion?

What We Learned

Building CareLens AI reinforced an important lesson:

AI should not replace human judgment—it should strengthen it.

The most valuable feature was not the model itself, but the ability to explain recommendations, quantify uncertainty, and keep humans in control.

We learned that trustworthy AI requires more than predictions.

It requires evidence.

What's Next

Future enhancements include:

  • Population-adjusted healthcare access scoring
  • Travel-time analysis and routing
  • Continuous learning from user feedback
  • Integration with public health datasets
  • Multilingual support for regional users

Conclusion

CareLens AI helps healthcare planners make better decisions by transforming messy healthcare data into evidence-backed insights.

Because when resources are limited and lives are affected, decisions should be driven by trust—not assumptions.

Built With

  • ai/bidashboard
  • databricks
  • genie
  • lakebase
  • lakehouse
  • lakehouseapp
  • mlflow
  • pyspark
  • unitycatalog
Share this project:

Updates