Reduce the coding backlog without removing coder review

What this product is, exactly

For Riverside’s outpatient ortho pilot, the engine is a pre-coder and a documentation checker. It recommends ICD-10 and CPT, highlights the source sentences, applies local payer edits as warnings, and builds a billing package that cannot leave the HIM queue until a certified coder accepts it.

It is not an autonomous coder. It is not a claims-submission bot. If a health system later wants query drafting or inpatient DRG grouping, those are separate workflows with their own thresholds.

What enters the system

1. 01

   The chart

   Operative notes, E/M notes, orders, and prior codes from Epic for the encounter.

2. 02

   Rules

   Code sets, the charge master, and payer edits the HIM team already maintains — not a generic internet codebook.

3. 03

   History

   Months of this facility’s coded encounters, rejects, and coder comments, used as the test the next model must beat.

What happens overnight

1. 01

   Read and cite

   Diagnoses and procedures are proposed only with a pointer into the note. No citation, no recommendation.

2. 02

   Thresholds

   Lines at or above the configured confidence — 0.85 in this walkthrough — can be packaged for one-click coder accept. Lines below that, or designated modifiers, always stop.

3. 03

   Documentation gaps

   Missing laterality in a header, absent indication, or a same-day E/M without a separately identifiable service become coder questions, not billed guesses.

4. 04

   Package, do not submit

   Accepted packages can be written to the charge workqueue after the coder’s accept. The engine never sends the claim.

What the coder sees in the morning

James Wu has 118 packages waiting for a single accept and 27 exceptions. ENC-88291 is an exception because modifier 25 is 0.61. He can read paragraph 1 and paragraph 3 next to the recommended codes, see the Blue Cross same-day edit, and decide whether the E/M is separately identifiable.

His accept is the legal coding act. His reject, with a reason, is the training event.

Automation boundaries

How a correction becomes a validated improvement

If James drops modifier 25, the encounter becomes a negative example for that modifier given this note structure. If he keeps it, the opposite.

A hypothesis that always attaches 25 to 20611 fails Riverside history and is discarded. Sentence-level citations and laterality highlighting pass. The winning model changes what is pre-coded confidently — not what is billed without a coder.

What a 30-day pilot measures

One specialty. No denial-rate claims until coding quality is compared to the baseline audit sample.

Implementation

Systems

Epic or equivalent EHR, encoder, charge master, payer-edit library.

Data required

Coded encounters for the specialty, audit findings, and the current query templates.

Deployment

Hardware in the health-system data center or HIM VLAN. Shadow on historical charts, then supervised pre-code, then workqueue writes after accept.

Who operates it

Certified coders accept every package. HIM owns thresholds. Compliance owns the audit sample. We do not access charts remotely.

Security and deployment architecture

Clinical notes are not a context window for an external model. Inference, logs, and weight updates stay on the hospital-controlled machine. Off-box telemetry is hardware health only.

The architecture brief covers network placement, encryption at rest, and how a support visit works without copying charts off-site.