Clinical AI Practice

Automated medical coding in production

— what works, and where most systems fall short

AI has significantly advanced automated CPT, modifier, and ICD-10 code mapping. Achieving production-scale performance now depends primarily on clinical ontology depth, documentation variability, and seamless revenue cycle integration.

Production benchmarks
95%+
F1 on discharge summaries
60%
Reduction in coder review time
4 wks
Typical integration to go-live
Zero
EHR workflow changes required
The Production Gap

Where automated coding
systems consistently fall short

Systems that perform well in controlled evaluations degrade against the conditions that define real hospital environments: dictated notes, multi-procedure encounters, and payer-specific rules that exist outside any standard code set.

Dictation artefacts and non-standard abbreviations

Production notes contain dictation errors, unconventional abbreviations, and fragmented sentence structure that standard NLP pipelines are not trained to handle. "Pt c/o SOB x3d, r/o PE, started on LMWH" is representative.

Pt c/o SOB x3dr/o PE, started LMWH??? ??? SOB ??? ???
1 of 5

"Accuracy on training data set and accuracy on production live datasets are often separated by 15 to 25 percentage points. The gap is addressable through re-training and implementing SOPs."

How the System Works

Vendor-Neutral Agentic Architecture that addresses High Accuracy and Operational Reliability

Each of the five stages corresponds to a specific failure mode above. PHI handling, ontology depth, assertion status, and payer rule integration are structural properties of the pipeline sequence — not features added on top. Click any stage to explore it.

Ingest
EHR notes · PDFs · dictations
De-identify
PHI removed before model sees text
Extract
Clinical NLP · negation · NER
Code
ICD-10 · CPT · payer rules applied
Review
Accept · modify · reject with source text
Stage 01 — Ingestion

Document ingestion — any format, any delivery method

The pipeline accepts clinical documentation in every format hospitals actually produce — dictated notes, scanned PDFs, structured EHR exports, and HL7 feeds. Format and delivery method are decoupled from downstream accuracy. An audit record is created at the point of ingestion, before any processing begins.

  • HL7 v2 and FHIR API connectors

    Direct integration with Epic, Cerner, and Meditech. Triggers IT security review — plan for 4–8 additional weeks. Not the fastest initial path.

  • SFTP batch ingestion

    The fastest deployment path. Notes are exported from the EHR and transferred via encrypted file exchange. Bypasses the IT security review process entirely. Most clients start here and migrate to direct API later.

  • OCR for scanned documents

    Scanned PDFs processed through an optical recognition layer before NLP. Recognition quality is a direct input to downstream extraction accuracy — poor scan quality is surfaced at ingestion, not discovered at coding.

  • Ingestion audit log

    Every document receives a unique ingestion record: timestamp, source system, format, and document type. Audit trail begins at entry — before de-identification, before any model sees the text.

Ingestion methods — trade-offs
METHODDEPLOYMENTSFTP / BatchEncrypted file transferFASTESTHL7 v2 / FHIRDirect EHR connectorIT REVIEW REQ.PDF / ScannedOCR preprocessing layerSCAN QUALITY ↑Direct APIReal-time single-noteIT REVIEW REQ.

SFTP is the recommended starting point for most deployments. IT security review for direct API access typically adds 4–8 weeks — plan for it early or start with SFTP.

Human-in-the-loop design

Workflow re-imagined to incorporate efficient & fast human-in-the-loop review and audits

Accurate code extraction is necessary, but not sufficient. The review interface is what converts accuracy into measurable time savings.

Coder review queue — interface model
CODING QUEUE4 PENDINGDischarge summary · [PHI redacted] · 03/28/2026E11.9T2DM without complications"Hx of T2DM" — line 297%ACCEPTR06.00Dyspnoea, unspecified"SOB x3d" — line 394%ACCEPTZ79.4Long-term insulin useLow confidence — verify source before accepting74%REVIEWI48.91Atrial fibrillation, unspec."parox. AF" — line 788%ACCEPT3 accepted · 1 pending reviewSubmission authority: coding teamSUBMIT TO BILLING

Amber row = low confidence. Final submission authority stays with the coding team.

10–15 min
Per chart
manual coding
→
< 4 min
Per chart
AI-assisted review

Coder authority over final submission is preserved throughout. The time per decision collapses.

Source-text evidence
Every suggestion links to the exact clinical text that generated it — validation in seconds, not minutes.
Confidence scoring
High-confidence codes accept in one step. Low-confidence flags surface only the cases that need judgment.
One-step review
Accept, modify, or reject per suggestion without leaving the queue.
Feedback loop
Every coder rejection is a labelled training signal. Accuracy compounds against your own documentation patterns.
Integration & Deployment

Integration considerations
and deployment timeline

The most common source of extended deployment timelines is not technical complexity — it is the security review process triggered by direct EHR API access. Understanding this early significantly affects project planning.

Exhibit 6 — Typical deployment timeline: integration to first production output
KickoffWeek 2Week 3Week 4ProductionDISCOVERYSOP alignment, payer mappingPIPELINE SETUPPHI validation, integrationPILOT — 500 CHARTSAccuracy measurement on your dataPRODUCTIONFull rollout + feedback loopSFTP path — no IT security review triggered, typically 2 weeks fasterDirect EHR API — triggers full IT security review, add 4–8 weeks

Integration path selection is the single largest determinant of go-live timeline. Secure file exchange (SFTP) bypasses the IT security review process that direct API access triggers.

Connection Options

  • REST API — structured JSON output, most common
  • Secure SFTP — fastest to deploy, no security review triggered
  • Direct EHR connector — Epic, Cerner, Meditech
  • RCM system feed — Optum, Waystar, Experian Health
Accuracy

The benchmark standard for measuring accuracy for operational reliability

F1 score — the harmonic mean of precision and recall — is the standard production measure. The OIG compliance threshold is ≥95%. Aggregate vendor figures routinely conceal the variance by document type and specialty that is most material to procurement decisions.

Precision
Of all codes assigned, what proportion were correct. High precision means a low false-positive rate — fewer incorrect codes submitted to payers.
Recall
Of all codes that should have been assigned, what proportion were captured. High recall means a low missed-code rate and lower revenue leakage.
F1 Score
Harmonic mean of precision and recall. The standard production metric — a single figure that penalises both incorrect codes and missed codes equally.
OIG Standard
The Office of Inspector General defines ≥95% accuracy as the compliance threshold for hospital coding programs subject to external audit.

Specialties by AI coding readiness

AI coding maturity varies by specialty — determined by documentation structure, code set depth, and payer rule density. The specialties below are grouped by deployment readiness, from production-proven to advanced-tuning required.

Production-readyDeployable today — high accuracy, widely validated
E&M Outpatient
Primary Care · Preventive Care · Internal Medicine · Paediatrics · Family Medicine
High-volume, structured encounter notes. Consistent documentation patterns drive strong baseline F1. Ideal first deployment for health systems.
HCC Risk Adjustment
Medicare Advantage · CMS-HCC · RAPS · EDPS
Hierarchical Condition Category coding for risk-adjusted payment models. AI excels at surfacing chronic conditions documented but not coded — directly improving RAF scores.
Radiology
Diagnostic Imaging · Interventional · Nuclear Medicine
Structured radiology reports provide highly consistent NLP inputs. Among the highest production F1 rates across all specialties.
Pathology & Labs
Clinical Pathology · Anatomic Pathology · Molecular Diagnostics
Structured reporting formats and well-defined code sets yield reliable automation with minimal post-processing.
Strong performance with tuningProduction-deployed with specialty-specific model adaptation
Inpatient Medicine
Hospital Medicine · DRG Assignment · Discharge Summaries · MS-DRG
Multi-diagnosis inpatient encounters require accurate principal diagnosis sequencing under UHDDS rules. DRG impact makes sequencing accuracy as important as code accuracy.
Emergency Medicine
ED E&M · Critical Care · Observation · Facility Coding
High rule-out documentation density makes negation handling critical. Specialty-tuned assertion models deliver strong results.
Surgery
General Surgery · Orthopaedics · Neurosurgery · Gynaecology · ENT
Operative note coding requires laterality, approach, and multi-procedure handling. Single-procedure notes perform well; multi-procedure cases benefit from tuned extraction.
Behavioural Health
Psychiatry · Psychology · Substance Use · Crisis Care
Narrative-heavy documentation with E&M and psychotherapy code interaction. Tuned models handle assertion status in psychiatric notation reliably.
Advanced — specialty models requiredHighest code set depth and payer rule density
Cardiology
Interventional · Electrophysiology · Heart Failure · Cath Lab
Dense procedure code sets, complex modifier interactions, and payer-specific bundling rules require a dedicated cardiology extraction layer.
Oncology
Medical Oncology · Radiation · Haematology · Chemotherapy Administration
Chemotherapy administration codes, staging, and treatment sequencing require deep ontology coverage and continuous model updates as protocols evolve.
Standards & Compliance

Code systems and compliance
architecture

The system outputs to the standard code sets and interoperability formats in use across hospital billing and RCM operations. PHI handling is architecturally integrated, not a configuration option.

ICD-10-CMICD-10-PCSCPTHCPCS Level IISNOMED CTRxNormLOINCFHIR R4HL7 CCDHIPAA — PHI De-identificationSOC 2 Available
Knowledge Base

Our extensive knowledge base

Technical depth for data and engineering teams, operational guidance for RCM leadership, and a coding practice track for coders adapting to AI-assisted workflows. Benchmark data and case studies are included within the relevant track.

Infrastructure and modelling decisions that determine production performance — including benchmark data on accuracy by document type and specialty.

10 topics
  • T.01Foundational constraint
    Why 7-character ICD-10 specificity is a medical ontology problem, not a model training problem
    Language models reliably identify the diagnostic category. Laterality, encounter type, and episode of care require live traversal of the ICD-10 hierarchy — a constraint no training corpus eliminates.
    ICD-10-CMUMLSOntology
    Read the article →
    −
  • T.02Technical depth
    Clinical NLP on real documentation — dictation artefacts, abbreviations, and fragmented syntax
    +
  • T.03Foundational constraint
    Negation detection, uncertainty modelling, and assertion status — why standard NER is insufficient
    +
  • T.04Technical depth
    PHI de-identification at the pipeline level — why sequence matters and where most architectures fail
    +
  • T.05Technical depth
    Fine-tuned clinical models versus general-purpose LLMs — where the performance gap materialises
    +
  • T.06Foundational constraint
    Measuring coding model accuracy correctly — what vendor figures typically measure vs what organisations need
    +
  • T.07Technical depth
    Multi-procedure code extraction and sequencing — the encounter type that exposes depth limits
    +
  • T.08Technical depth
    Payer rule integration — why code correctness and claim submission correctness are different problems
    +
  • T.09Benchmark Data
    ICD-10/CPT accuracy benchmarks — production F1 by document type (2024 cohort)
    +
  • T.10Benchmark Data
    Specialty-stratified accuracy benchmarks — F1 by specialty and document type
    +
Contact

Talk to someone who has built this in production

Most conversations start with a specific technical question — not a sales process. If you have a coding or NLP problem, we can tell you quickly whether what we do fits.

Request a BriefView Solutions →