BRENNEN

Industries / Health / Healthcare / Medical billing

Coding

Turning the record of what was done for a patient into standard codes for diagnoses and procedures, and grouping those codes into the categories that set the price.

6

Milestones achieved, from 1966 to 2024.

2

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 1966

    A list of procedure codes

    The American Medical Association (AMA) published the first edition of Current Procedural Terminology (CPT), a set of standard terms, later five-digit codes, for the procedures doctors perform. CMS, the US agency that runs Medicare and Medicaid, now treats CPT as Level I of its national coding system for claims, and an independent AMA panel updates it three times a year. When the US made national billing standards in 2000, it named CPT the required code set for physician services even though it is the AMA's own product: royalties for electronic use were then $10 per user, while the government's own code sets could be downloaded free.

    Lessonthe national language of US doctors' bills is privately owned, so anyone who builds billing software or tests pays for a license.

    ama-assn.orggovinfo.gov

    Credit

    American Medical Association

  2. 1983

    Paying by patient group

    A diagnosis-related group (DRG) puts each hospital stay into one of a few hundred classes of patients expected to use similar resources, based on the diagnosis and procedure codes. By CMS's account, work on DRGs began at Yale University in the late 1960s, New Jersey first used them to pay hospitals a fixed amount per stay in the late 1970s, and in 1983 Congress made DRGs the basis of a national prospective payment system for Medicare hospital stays, meaning the price is set before the stay rather than repaid from costs afterward.

    cms.gov

    Credit

    Yale University School of Organization and Management; New Jersey State Department of Health; US Congress..pdf)

  3. 2004

    Germany prices stays by DRG

    After one optional year, Germany paid almost all hospital stays through its own G-DRG system from 2004. An evaluation for InEK, the hospitals' and funds' remuneration institute, found that from 2004 to 2006 the average stay fell from 7.77 to 7.47 days, while the case-mix index, the average weight of the coded cases, rose 2.0% a year; the authors said this could come from changes in documentation and coding as well as from sicker patients. A later study of national hospital data for 2005 to 2012 found more diagnoses coded per record over time, which shifted the risk estimates researchers use to compare hospitals.

    Lessononce codes set the payment, coding becomes more complete, and that alone changes every statistic built on the codes.

    g-drg.depmc.ncbi.nlm.nih.gov

    Credit

    InEK; IGES Institut (evaluation, March 2010); Ulrike Nimptsch, Technische Universität Berlin (Health Services Research, 2016)

  4. 2012

    Coding accuracy measured in Britain

    A systematic review of 32 studies compared the codes in British hospital data with the patients' notes. The median accuracy was 83.2%, ranging from 50.5% to 97.8% between studies. Since Payment by Results began in 2002, which pays English hospitals per coded case, accuracy of the primary diagnosis rose from a median of 73.8% to 96.0%.

    Lessonwhen codes bring in money, hospitals invest in getting them right.

    pmc.ncbi.nlm.nih.gov

    Credit

    E. M. Burns, E. Rigby, R. Mamidanna and colleagues (Journal of Public Health, 2012)

  5. 2015

    The US switches to ICD-10

    On 1 October 2015 US health care moved from ICD-9 to ICD-10 codes for diagnoses (ICD-10-CM) and for inpatient procedures (ICD-10-PCS), years after most other countries had adopted ICD-10. The switch had been delayed twice from 2013, and organizations at an industry summit estimated the 2014 delay alone would cost them up to $5 million to $10 million each. In a time study before the switch, coders took 17.71 minutes, or 69%, longer per inpatient record with the new codes, and agreement between coders fell from substantial to moderate.

    Lessona new code set costs most in coders' time and in breaks in the data series, not in software.

    cms.govpmc.ncbi.nlm.nih.govpmc.ncbi.nlm.nih.gov

    Credit

    CMS and the Centers for Disease Control and Prevention; Mary H. Stanfill (Perspectives in Health Information Management, 2014); Sheila V. Kusnoor and colleagues (JAMIA Open, 2019)

  6. 2024

    AI models miss billing codes

    Researchers gave GPT-3.5, GPT-4, Gemini Pro and Llama 2 the descriptions of billing codes used over 12 months in the Mount Sinai Health System and asked for the codes. The best, GPT-4, returned the exact code for 33.9% of ICD-10-CM codes, 45.9% of ICD-9-CM codes and 49.8% of CPT codes, and the models often produced codes that were imprecise or did not exist.

    Lessona model that writes fluent text still has to look codes up rather than recall them.

    ebi.ac.uk

    Credit

    Ali Soroush, Eyal Klang and colleagues (NEJM AI, 2024)

Watch for

  1. Hospital stays coded without people

    AKASA says its AI now codes complex inpatient stays and checks clinical documentation with no human step, in under 90 seconds after discharge where a coder takes 30 to 60 minutes. It says third-party blinded studies found its AI matched or beat human coders on assigning the Medicare DRG, the principal diagnosis and other measures, and that its customers account for about 10% of US inpatient discharges.

    Achieved whenthe evaluations, with their samples, the outside party's name and error rates, are published, or a hospital's auditors report results on cases coded only by the AI.

    akasa.com

    Claimed by

    AKASA, October 2026

  2. Clinic visits coded automatically

    Your Health, a group serving older patients in US clinics, assisted living and nursing homes, says Fathom's AI has coded 95.5% of its visits without human review since October 2025, that coding accuracy rose from 96.3% under human coders to 98.3%, and that its average risk adjustment factor, the score that sets some Medicare payments, rose 0.66 in a year. These are the two companies' figures, as reported by HIT Consultant.

    Achieved whenan outside coding audit, or a payer's review of the medical records, confirms the accuracy rate and finds the higher scores supported.

    hitconsultant.net

    Claimed by

    Your Health and Fathom, March 2026

Open

  1. 01

    An answer key for DRGs

    A public set of 500 invented hospital stays, each with a discharge summary, an operation note where relevant and key test results, written by clinicians and coders for people who never existed. Two certified inpatient coders code each stay separately in ICD-10-CM and ICD-10-PCS, settle disagreements, and run the agreed codes through Medicare's grouper to get the DRG. Then score AI coding tools and general AI models on the principal diagnosis, the procedure codes and the DRG, and count what each wrong DRG would cost or add at Medicare's public payment weights. Inpatient hospital claims use ICD-10-PCS for procedures, so the set needs no CPT license.

    The brief
    Credit · unclaimed
  2. 02

    Coding straight into ICD-11

    A test of coding in ICD-11, the World Health Organization's newest disease classification, on 300 invented case summaries whose right answer needs postcoordination, meaning a main stem code joined to extension codes for details such as severity or body site. Two coders trained in ICD-11 code each case with WHO's coding tool and agree an answer. Then score AI models with and without access to WHO's ICD-11 programming interface, and time human coders on the same cases.

    The brief
    Credit · unclaimed

Open 01 · the brief

An answer key for DRGs

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A public set of 500 invented hospital stays, each with a discharge summary, an operation note where relevant and key test results, written by clinicians and coders for people who never existed. Two certified inpatient coders code each stay separately in ICD-10-CM and ICD-10-PCS, settle disagreements, and run the agreed codes through Medicare's grouper to get the DRG. Then score AI coding tools and general AI models on the principal diagnosis, the procedure codes and the DRG, and count what each wrong DRG would cost or add at Medicare's public payment weights. Inpatient hospital claims use ICD-10-PCS for procedures, so the set needs no CPT license.

Why it is still open

DRGs have been predicted from notes before, but only from real records that cannot be shared. DRG-LLaMA, a LLaMA model tuned on 236,192 discharge summaries from MIMIC-IV, a research database of real patients' records, named the right DRG as its first choice in 52.0% of stays, by its authors' count; only credentialed researchers can get the notes, and no coder labeled them for the test. The 2024 Mount Sinai test scored code lookup from short descriptions, not whole charts. Falis and colleagues had GPT-3.5 write discharge summaries from lists of codes taken from MIMIC-IV and found GPT-3.5 unsuitable for coding; they did not test DRGs. AKASA's blinded evaluations are not public. A search in November 2026 found no public test that scores DRG assignment.

pmc.ncbi.nlm.nih.govarxiv.org

Built with what exists

CMS's MS-DRG grouper, published with its Java source code; the free ICD-10-CM and ICD-10-PCS code files and coding guidelines from CMS; Synthea, an open-source generator of invented patient histories, as a starting point for the stays; general AI models through their public interfaces and open-weight models run locally; a short Python scoring script.

cms.gov

  1. First step

    Write 20 stays in one DRG family, such as heart failure with and without major complications, have two coders code them separately, and run both sets through the grouper.

  2. Done looks like

    500 stays covering at least 20 of the 25 major diagnostic categories, agreement between the two coders reported before reconciliation, at least four tools scored on DRG match rate and net payment error per 100 stays, and the set published with a DOI under an open license.

Open 02 · the brief

Coding straight into ICD-11

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A test of coding in ICD-11, the World Health Organization's newest disease classification, on 300 invented case summaries whose right answer needs postcoordination, meaning a main stem code joined to extension codes for details such as severity or body site. Two coders trained in ICD-11 code each case with WHO's coding tool and agree an answer. Then score AI models with and without access to WHO's ICD-11 programming interface, and time human coders on the same cases.

Why it is still open

The US has not adopted ICD-11. Researchers at the National Library of Medicine and the National Center for Health Statistics found that ICD-11 stem codes alone fully covered 35.2% of 1,725 ICD-10-CM codes, and 89.4% with postcoordination, so postcoordination is where a US move would succeed or fail. The nearest AI test, MAX-EVAL-11, uses 10,000 real discharge summaries from the MIMIC-III research database with mapped ICD-11 codes; it found exact-match rates of 0 to 4.8% and does not report postcoordination. A June 2026 six-language set scores 300 code titles, not cases. A search in November 2026 found no public set of cases coded natively in ICD-11 with postcoordination.

ebi.ac.ukaclanthology.org

Built with what exists

WHO's ICD-11 browser, coding tool and programming interface, and its ICD-10 to ICD-11 mapping tables; ICD-11 content is licensed CC BY-ND 3.0 IGO, so the set publishes codes and links rather than altered WHO text; general AI models; a spreadsheet for the two coders' answers.

who.intzenodo.org

  1. First step

    Take 20 common ICD-10-CM codes that the 2023 study found need postcoordination, write one invented case for each, and have two coders code them separately with WHO's coding tool.

  2. Done looks like

    300 cases, at least 100 of which need two or more extension codes, agreement between the coders reported, and for each of at least four models the share of cases where the whole code cluster matches, with and without the interface, published with a DOI.