AI Scribe for GI: Managing the Prague Classification for Barrett's Esophagus

How Scribing.io Captures Prague C & M as Structured Data to Eliminate Surveillance EGD Denials

• The Documentation Gap Every GI Practice Ignores

• Clinical Logic: From 7 Denials/Month to <1%

• Step-by-Step Logic Breakdown: How the Hard-Stop Works

• Technical Reference: ICD-10 Documentation Standards

• Seattle Protocol Biopsy Mapping as Structured Data

• The Medical Necessity Pack: One-Click Pre-Auth Packaging

• EHR Integration Architecture: Epic, Cerner, athenahealth

• Cross-Specialty Documentation Patterns

• Book a 15-Minute Workflow Audit

The Documentation Gap Every GI Practice Ignores: Barrett's Is a Procedural Dataset, Not a Symptom Narrative

Seven denied surveillance EGDs last month. $24,320 in held revenue. Ninety minutes of staff time per case spent chasing addenda, calling outside offices, and resubmitting documentation that should have been clean on the first pass. The physicians assumed their scribe captured it. The scribe—an ambient AI platform marketed as "GI-optimized"—logged "Barrett's esophagus" in the assessment and moved on. No Prague C. No Prague M. No centimeter values. No link to the prior encounter's measurements. The payer's utilization management nurse searched the note for structured evidence of disease extent, found a narrative paragraph, and stamped the authorization request denied.

This is not an edge case. It is the structural failure mode of every AI scribe that treats Barrett's esophagus as a condition to be coded rather than a procedural dataset to be captured. Scribing.io was built to close this gap—specifically, to treat Prague C & M as discrete, unit-validated, EHR-structured observations that carry forward across encounters and auto-package into payer-ready documentation. The difference is not incremental. It is categorical: the difference between a note that generates a clean claim and a note that generates a denial.

A close read of the current competitive landscape—including the leading vendor listicles comparing AI scribes for gastroenterology—reveals a consistent omission: not a single platform addresses the Prague Classification as a discrete, structured documentation event. Vendors discuss polyp morphology, MELD scores, and variceal grading as markers of specialty depth. Those matter. But none of them explain how Barrett's-specific documentation failures cascade into surveillance EGD denials, how Seattle-protocol biopsy compliance should be captured at the point of care, or how the absence of a native Prague C & M field in Epic, Cerner, or athenahealth creates a systemic documentation gap that payers exploit.

The 2022 AGA Clinical Practice Update on Barrett's Esophagus is explicit: surveillance intervals are segment-length dependent, determined by Prague C & M measurements and dysplasia status. The ASGE quality indicators for upper endoscopy list Prague classification documentation as a performance metric. Payers read these guidelines. Their utilization review criteria reflect them. When your note lacks the data these guidelines require, the payer has a guideline-backed reason to deny.

We have documented similar structured-data capture patterns across other specialties. In Cardiology, Scribing.io captures ejection fraction, valve gradients, and NYHA class as discrete fields—not narrative mentions—because those values drive prior-auth decisions for imaging and interventional procedures. In Psychiatry, PHQ-9 and GAD-7 scores are captured as validated integers linked to treatment-plan logic. The principle is identical: if a data element drives a payer decision, it must exist as structured, queryable, auditable data in the EHR—not as a string buried in paragraph five of a free-text note.

Scribing.io Clinical Logic: From 7 Denials/Month to <1% — A Before-and-After Workflow Analysis

The following scenario is modeled on operational patterns reported by multi-physician GI groups performing high-volume surveillance endoscopy.

Before Scribing.io

A 6-physician GI group averages 38 surveillance EGDs per month. Last month, 7 were denied because the prior operative note lacked discrete Prague C & M scores. Each denial triggered:

• Average delay: 21 days from initial submission to resolution

• Held revenue: $24,320 across the 7 cases

• Staff time per case: ~90 minutes chasing outside records, requesting physician addenda, resubmitting clinical documentation, fielding payer callbacks

• Total monthly staff burden: ~40 hours of combined upstream rework and denial management (pulling old notes, calling referring offices, drafting appeal letters, re-keying data into payer portals)

• Clinical risk: Delayed surveillance intervals for patients with known Barrett's—a compliance exposure under CMS quality metrics and a medicolegal liability if interval cancer develops during the delay

Root cause analysis: Clinicians verbalized Prague values during the procedure. The generic AI logged "Barrett's esophagus" as a condition in the assessment—correct as a diagnostic label, useless as a payer-facing dataset. No C value. No M value. No centimeter units. No carry-forward from prior encounters. No link between the Prague measurement and the biopsy protocol. The payer's UM nurse saw ICD-10 code K22.70 but no structured evidence of disease extent or interval change. Authorization denied per the payer's internal application of AGA surveillance interval guidelines.

After Scribing.io

Step-by-Step Logic Breakdown: How Scribing.io's Hard-Stop Solves the Prague Problem

The anchor truth: Payers deny surveillance EGDs if the prior note lacks Prague C & M scores. Generic AI treats Barrett's as a "symptom." Scribing.io treats it as a "procedural dataset" required for future billing approval. Here is the granular, step-by-step logic of how that treatment translates into a solved problem.

Step 1: Ambient Capture with Prague-Specific NLP Triggers

During the procedure, the endoscopist dictates findings. Scribing.io's GI-tuned language model listens for Barrett's-related utterances: "circumferential extent," "maximum extent," "Prague," "C and M," "tongues," "salmon-colored mucosa," and numeric centimeter values spoken in proximity to GEJ landmarks. Unlike general-purpose ambient AI that tokenizes these as narrative strings, Scribing.io's model extracts two discrete numeric values (C and M) and one landmark value (GEJ distance from incisors in cm) and stages them for structured write-back.

Step 2: Hard-Stop Validation Gate

The note enters a validation state before finalization. If a Barrett's-related diagnosis or finding is detected in the encounter and Prague C, Prague M, or GEJ landmark fields are empty, the note cannot be signed. The clinician sees a modal prompt: "Prague C & M required for Barrett's documentation. C: ___ cm M: ___ cm GEJ: ___ cm from incisors." Numeric range validation (0–20 cm for C and M; 20–45 cm for GEJ from incisors) prevents data-entry errors—a C value of "34" (a GEJ distance accidentally entered in the wrong field) is rejected with an inline explanation.

This hard-stop is the single most consequential design decision in the playbook. It converts Prague documentation from a clinician-dependent, memory-dependent, scribe-dependent act into a system-enforced requirement. The JAMA evidence on structured documentation improving data completeness is well-established; Scribing.io applies that principle to a specific, high-stakes GI use case.

Step 3: Discrete Data Write-Back to the EHR

Once validated, Prague C, Prague M, and GEJ landmark are written to the EHR as discrete observation fields—not as tokens in a free-text note. In Epic, these map to SmartData Elements (SDEs) or flowsheet rows. In Cerner (Oracle Health), they map to discrete result values within a procedure event. In athenahealth, they map to structured clinical data fields via the athenahealth API. The critical property: these values are queryable by reporting tools, auditable by compliance teams, and extractable by payer-facing documentation workflows without human parsing of narrative text.

Step 4: Prior-Encounter Prague Carry-Forward and Delta Calculation

When a patient with a documented Barrett's history presents for surveillance, Scribing.io queries the EHR's discrete data store for the most recent Prague C & M values and their encounter date. These are auto-populated into a "Prior Barrett's Status" section of today's note:

• Prior Prague (2024-09-14): C2 M4

• Current Prague (2026-01-22): C2 M5

• Delta: C stable, M +1 cm

This carry-forward serves two functions. Clinically, it gives the endoscopist an instant reference for interval change without toggling between encounters. Operationally, it provides the payer with the exact data pair they need to validate that the surveillance interval is guideline-concordant and that the procedure is medically necessary. The delta calculation is the specific data element that UM nurses search for and cannot find in generic AI-generated notes.

Step 5: ICD-10 Assignment Driven by Prague + Pathology

Scribing.io does not assign a Barrett's ICD-10 code based on the diagnosis label alone. The code is determined by the intersection of endoscopic findings (Prague C & M confirming columnar-lined esophagus) and pathology results (dysplasia status). The system holds the code at K22.719 (unspecified dysplasia) until pathology results are received and linked. Once pathology posts, the code is refined to K22.70 (no dysplasia), K22.710 (low grade), or K22.711 (high grade). This prevents the common error of finalizing a claim with an unspecified code when a specific code is available—a pattern that triggers CMS audits and payer downcoding.

Step 6: CPT Mapping with Structured Justification

The encounter's CPT code (43235 for diagnostic EGD; 43239 for EGD with biopsy) is linked to the structured Prague data and Seattle-protocol biopsy documentation. The justification payload—not a narrative letter, but a structured data bundle—includes: indication (surveillance of known Barrett's, ICD-10 code), prior Prague with date, current Prague, interval-change assessment, biopsy count and sites per Seattle protocol, and guideline citation. This payload travels with the claim as a machine-readable attachment.

Step 7: Medical Necessity Pack Generation

For payers requiring prior authorization, or when attaching clinical documentation to a claim, the front-desk or billing team clicks one button: "Generate Medical Necessity Pack." The system assembles:

1. Prior encounter Prague C & M with date

2. Current encounter Prague C & M with date

3. Delta calculation

4. Seattle-protocol biopsy compliance summary

5. Pathology result (linked or pending)

6. ICD-10 code with specificity rationale

7. CPT code with procedure justification

8. AGA guideline–concordant surveillance interval citation

This pack is exported as a PDF for fax-based payers or as a structured data file (C-CDA or X12 835/837 attachment) for electronic submission. The ~90 minutes per denied case that staff previously spent assembling this information manually drops to under 2 minutes for generation and transmission.

Technical Reference: ICD-10 Documentation Standards for Barrett's Esophagus

Accurate ICD-10 assignment for Barrett's esophagus is a documentation problem, not a coding-department problem. The dysplasia status that determines the correct code is established by pathology, but the documentation linkage between endoscopic findings and histologic results must be explicit and structured in the encounter record for the code to survive audit.

How Scribing.io Ensures Maximum Code Specificity

The most common coding failure in Barrett's documentation is premature finalization: the claim goes out with K22.719 (unspecified dysplasia) because the note was signed before pathology resulted, and no one updated the code afterward. Per AMA CPT/ICD-10 documentation standards, the highest specificity code supported by available clinical evidence must be used at the time of claim submission.

Scribing.io addresses this with a pathology-linkage workflow:

1. At encounter close, if biopsies were taken and pathology is pending, the ICD-10 field is set to K22.719 with an automated flag: "Awaiting pathology—code refinement required."

2. When pathology results post to the EHR (via HL7 ORU message or manual entry), Scribing.io surfaces a reconciliation task to the signing clinician: "Pathology for [Patient] received. Barrett's dysplasia status: [No dysplasia / LGD / HGD]. Confirm code update: K22.719 → K22.70 / K22.710 / K22.711."

3. Upon clinician confirmation, the encounter's ICD-10 is updated, the claim is released (or amended if already submitted), and the Medical Necessity Pack is regenerated with the final code and pathology summary.

This three-step loop eliminates the unspecified-code problem at its source. It also creates an auditable trail demonstrating that the practice systematically refines codes based on pathology—a compliance posture that protects against OIG upcoding investigations and payer recoupment actions targeting nonspecific coding patterns.

Seattle Protocol Biopsy Mapping as Structured Data

The Seattle protocol (4-quadrant biopsies every 1–2 cm of Barrett's segment) is the standard-of-care biopsy methodology for Barrett's surveillance. Adherence is a quality metric, a medicolegal standard, and—increasingly—a payer checkpoint. If a UM reviewer sees a Prague M of 5 cm but only 3 biopsy sites documented, the note suggests protocol non-compliance, and the reviewer may question whether the procedure was performed to guideline standards.

Scribing.io captures biopsy sites as a structured map:

If the expected number of biopsy levels (based on Prague M and the q2cm interval) exceeds the documented biopsy sites, the system flags the gap in real time: "Seattle protocol gap: Expected biopsies at 4 levels (q2cm × M5); 3 levels documented. Confirm additional biopsy site or document clinical rationale for deviation." This real-time flagging prevents documentation gaps that weaken both clinical quality and claim defensibility.

The Medical Necessity Pack: Anatomy of a One-Click Pre-Auth Package

The Medical Necessity Pack is not a letter. It is a structured clinical data package assembled programmatically from discrete EHR fields. Its components:

1. Patient demographics and insurance identifiers — pulled from the EHR registration module

2. Prior Barrett's encounter summary: Date, Prague C & M (discrete values), pathology result, ICD-10 code assigned

3. Current encounter summary: Date, Prague C & M (discrete values), Seattle-protocol compliance status, biopsy count and sites

4. Interval change assessment: C delta, M delta, dysplasia status change (if pathology available)

5. Guideline-concordant surveillance interval citation: Auto-generated based on prior dysplasia status and segment length, citing AGA 2022 Barrett's guidelines

6. ICD-10 code with specificity rationale: Why K22.70 vs. K22.710 vs. K22.711, supported by pathology linkage

7. CPT code with procedure justification: 43235 or 43239, with structured indication

8. Ordering physician attestation: Digital signature block

Total generation time: under 2 minutes. Contrast with the manual process: staff opens the prior encounter, scans free text for Prague values (often unsuccessfully), calls the physician for an addendum, waits for the addendum, re-opens the payer portal, attaches documents, and submits. That process averages 90 minutes per case and fails frequently because the addendum itself is often another free-text paragraph that the payer's UM nurse still cannot parse efficiently.

EHR Integration Architecture: Epic, Cerner, athenahealth

The absence of native Prague C & M fields in major EHR platforms is the root infrastructure problem. Scribing.io does not work around this gap—it fills it.

Deployment typically requires 2–4 weeks of configuration, including SDE or field creation, validation rule setup, carry-forward logic activation, and Medical Necessity Pack template customization. Scribing.io's implementation team handles the EHR-side build in coordination with your IT and clinical informatics staff. No custom development is required from your internal team.

Cross-Specialty Documentation Patterns: Why Structured Clinical Data Capture Is Not a GI-Only Problem

The Prague C & M problem is a specific instance of a general pattern: specialty-critical measurements that drive payer decisions exist in most subspecialties, and most EHRs lack native structured fields for them. Scribing.io has addressed this pattern across multiple specialties:

• Cardiology: Ejection fraction, valve gradients (mean and peak), NYHA functional class, and LVEDD captured as discrete numeric values linked to imaging CPT codes and heart failure ICD-10 codes (I50.x). Prior-auth for advanced imaging (cardiac MRI, PET) requires these values in structured form.

• Psychiatry: PHQ-9, GAD-7, AUDIT-C, and Columbia Suicide Severity Rating Scale scores captured as validated integers with date stamps. Payers increasingly require score-based justification for medication changes and therapy authorization.

• Gastroenterology (this playbook): Prague C & M, Seattle-protocol biopsy mapping, MELD-Na for hepatology encounters, Forrest classification for ulcer bleeding—each captured as discrete, unit-validated data.

The architectural principle is consistent: identify the data elements that payers use to make authorization and payment decisions, ensure those elements are captured as structured observations (not narrative text), and package them for payer-facing workflows automatically. This is what separates a documentation tool from a revenue-cycle tool.

Book a 15-Minute Workflow Audit: Get Your Prague C & M Build Started This Week

If your practice performs surveillance EGDs for Barrett's patients, you are either capturing Prague C & M as structured data or you are generating preventable denials. There is no middle ground.

In a 15-minute Workflow Audit with Scribing.io's GI implementation team, you receive:

1. A Prague C & M discrete-field build spec for your EHR — whether you run Epic, Oracle Health (Cerner), or athenahealth, we deliver a ready-to-deploy SDE/field configuration document specific to your environment

2. A denial-risk score on your current EGD documentation template — we review your current procedure note template and identify every field that is missing, unstructured, or non-queryable for payer-facing documentation

3. A ready-to-send Medical Necessity Pack template — a pre-built package format that your billing team can use immediately to reduce surveillance EGD denials this month, even before full Scribing.io deployment

This is not a sales call. It is a technical consultation delivered by clinical informaticists who have deployed Prague-specific documentation workflows in GI groups ranging from 3 to 40 physicians. Book your 15-minute Workflow Audit at Scribing.io →

Surveillance EGD denials caused by missing Prague data are a solved problem. The question is whether your practice solves it this month or continues absorbing $24,000+ in monthly held revenue while staff spend 40 hours chasing documentation that should have been captured correctly the first time.