Reveleer: GenAI HEDIS evidence extraction that shows its source
HEDIS quality measures are central to how health plans are rated, and every measure has to be backed by evidence from the clinical record. At Reveleer, Clinical Quality is my domain, and I partnered with our AI team on a GenAI engine for HEDIS evidence extraction. The extracted evidence was 80% more accurate than the earlier version.
- My role
- Director of Product
- Company
- Reveleer
- Timeline
- Nov 2025 – Present
- Domain
- Clinical Quality
At a glance
The challenge
HEDIS abstraction means finding evidence for each measure across long clinical charts, with exclusions and date rules that are easy to get wrong.
What I led
Clinical Quality is my domain. With the AI team, a GenAI evidence-extraction engine built on a staged LLM workflow, deterministic dates, richer metadata and full traceability.
Result
Extracted HEDIS evidence 80% more accurate than the earlier version, improving abstractor accuracy and time to completion.
Context
In value-based care, health plans are measured on quality through HEDIS. For each measure, abstractors have to confirm from the chart whether care was delivered, when it happened, and whether an exclusion such as hospice or palliative care applies.
Any AI in that loop has to answer two questions every time: is this evidence really in the record, and can a reviewer see exactly where?
My role
I'm part of the core team shaping Reveleer's Unified Value-Based Care Platform initiative, where product domains, pods and roadmaps are organized around one end-to-end customer journey, and I own the Clinical Quality domain. I partnered with our AI team on LLM-based evidence extraction for HEDIS measures.
Product (me)
Clinical Quality domain owner, partnering with the AI team on the evidence-extraction engine.
AI team
Built the staged LLM workflow and its validations.
Abstractors
The end users, who verify the evidence they're shown.
How the engine works
Simplified view of the evidence-extraction workflow
- A staged LLM workflow instead of rules. A keyword pre-filter finds candidate pages, a fast, low-cost model strips out noise, and a reasoning model confirms the evidence. What survives passes measure-specific validations and exclusions such as hospice, palliative care and frailty. Each stage narrows the context, so the final model sees less noise.
- Deterministic dates. A rule-based module parses dates before the LLM sees them, so the model reasons over clean, structured dates instead of raw text.
- Richer metadata. Line-level provenance, multiple date types (collect, result and report dates) and care setting make rules possible that the legacy system couldn't reliably apply, such as excluding a blood pressure reading taken during an ER visit. Capturing the data and applying the full rule set are separate steps, and the second is still in progress.
- Traceability. Every piece of evidence links back to its source line, page, section and table cell, and a check catches line IDs the model made up. Abstractors can verify what they're shown, and hallucinated evidence gets caught.
Why these choices matter
1. Narrow the context at every stage
Cheap, fast stages remove noise before the reasoning model runs, so the most capable model works on the smallest, cleanest input.
Why it matters: accuracy improves when the final model sees less irrelevant text.
2. Keep dates out of the LLM's hands
Dates decide whether a measure is met, so they're parsed by rules before the model sees them.
Why it matters: the model reasons over dates it can't misread.
3. Evidence a reviewer can check
Every item links to its source, and invented line IDs are caught before an abstractor sees them.
Why it matters: in regulated work, a reviewer who can verify the source is the best safeguard against hallucination.
How it was measured
The engine was measured against the earlier version on the accuracy of the evidence it extracted.
| Metric | Before | After |
|---|---|---|
| HEDIS evidence accuracy | Earlier version | +80% |
| Workflow | Rules | 4-stage LLM workflow |
| Date handling | Read from raw text | Rule-based parsing first |
| Evidence provenance | — | Line, page, section, table cell |
Outcomes
- Extracted HEDIS evidence 80% more accurate than the earlier version.
- Better abstractor accuracy and time to completion, with evidence they can verify at the source.
- Hallucinated evidence caught by checking line IDs against the source document.
What I'd do next
Lessons
Give the most capable model the least noise. Narrowing the context at each stage is what lets the final model focus on the evidence.
Keep deterministic work deterministic. Dates and rules belong in code, not in a prompt.
Make evidence traceable. If a reviewer can't check it, it isn't evidence.
Building GenAI for a regulated market?
I work on GenAI where every piece of evidence has to trace back to its source. I'd welcome a conversation about yours.