Coding & NLP

Natural language processing reads the unstructured half of the chart that structured claims data never sees: progress notes, consult letters, discharge summaries, and problem lists. Those are where chronic conditions are described in prose but never coded to a diagnosis that reaches CMS.

The accuracy difference is measurable. Coder-only retrospective review typically lands at 75 to 80 percent first-pass accuracy. Invent Health's Coder Workbench delivers 85 to 90 percent first-pass accuracy using NLP alone, and up to 98 percent when paired with a human quality-assurance overlay. That is an 18 to 23 point improvement in first-pass accuracy over a coder-only workflow.

The mechanism matters as much as the number. NLP does not replace the coder. It reads every chart in the population rather than the sample a human team has time for, surfaces the suspected conditions and documentation gaps, and hands the coder a prioritized queue instead of a stack.

A documentation gap is a condition the clinical record supports but the submitted diagnosis does not, and finding one requires comparing the chart against the claim. Tools that only read claims cannot see them.

Invent Health's NLP-driven abstraction surfaces HCCs, suspected conditions, missed opportunities, and documentation gaps from unstructured chart data. AI-powered chart triage then auto-routes charts to coders based on complexity, documentation quality, and member risk impact, so the highest-value gaps are worked first rather than whatever surfaced alphabetically.

The output a compliance team should ask for is lineage: full traceability from chart to HCC capture to validation, supporting RADV, MAO-004, and internal coding audits. A gap you cannot trace back to its source document is a gap you cannot defend in an audit.

It holds up when every captured HCC carries its evidence. Invent Health maintains full lineage from source chart to HCC capture to validation, so each submitted diagnosis can be traced to the document and passage that supports it.

That trail is what a RADV audit actually tests. CMS now extrapolates audit findings across the contract, so an unsupported diagnosis is no longer a single-member recoupment. Audit-ready documentation, first, second, and third-level review workflows, and condition validation tracking are built into the coding workflow rather than assembled after an audit notice arrives.

Three lines of business in one workflow: CMS-HCC and RxHCC for Medicare Advantage, Medicaid models including CDPS and state-specific variants, and ACA HHS-HCC abstraction.

Running them in a single workflow matters for plans with dual-eligible or churned members, where the same member's history sits across Medicare and Medicaid and a single-model tool loses the longitudinal view that drives accurate suspecting.

Yes. The Coder Workbench is deployment-flexible by design: it runs with in-house coders, with Invent Health services, or with a third-party coding vendor or BPO. It is cloud-based and scalable.

Plans that already have a coding vendor typically use it as the NLP pre-pass and QA layer over that vendor's work, which is also the fastest way to benchmark whether the current vendor's first-pass accuracy is where the contract says it is.

Bring your own charts. Invent Health runs a 30-minute walkthrough where a plan brings a chart sample and leaves with an accuracy benchmark measured on its own population rather than a vendor demo dataset.

That is the right test. First-pass accuracy varies by population, documentation culture, and specialty mix, so a benchmark on your charts is the only number that predicts your result.