Industries / Health / Healthcare / HealthTech
Scheduling and billing
Scheduling and billing software runs the business side of care: it books appointments and referrals, registers patients, turns each visit or hospital stay into standard codes, sends claims to insurers or national payers, requests prior authorization (an insurer's approval before a test or treatment), and tracks denials and payments, a chain known as the revenue cycle.
7
Milestones achieved, from 1969 to 2024.
3
To watch for: announced, not yet confirmed outside.
3
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1969
Hospitals share a billing computer
IBM wrote the Shared Hospital Accounting System (SHAS) so that small and mid-size hospitals could share one IBM 360 mainframe for census, inpatient and outpatient billing, accounts receivable and the general ledger, sending punched-card data over leased 1,200-baud phone lines. In 1969 three former IBM Philadelphia staff raised $5 million from a Wall Street venture firm to buy their own 360 and sell, install and run the service as Shared Medical Systems (SMS); the company's filings say it was incorporated in January 1969. By its 1994 annual report SMS counted more than 2,000 customers, about 1,500 in North America and 500 in Europe, and revenue of $550.8 million.
Credit
IBM; Jim Macaleer, Harvey Wilson and Clyde Hyde (Shared Medical Systems)
1983
Paying hospitals by diagnosis
Diagnosis-related groups (DRGs) sort hospital stays into classes of patients whose diagnoses, procedures and personal details predict similar use of resources; Robert Fetter, John Thompson and colleagues at Yale published the classification in Medical Care in 1980. Medicare adopted DRGs for a prospective payment system, a price fixed in advance for each stay, effective 1 October 1983, so that, in the words of CMS's research journal, assigning a DRG code to a discharge record became "tantamount to the Medicare prospective payment"; grouper software makes that assignment from the codes on the record. The average Medicare stay fell from 9.8 days in 1983 to 8.7 days in 1985 by the same journal's figures. Twelve European countries and Australia built their own versions; Germany made one, adapted from Australia's, compulsory for acute hospitals in 2004, assigned by certified grouping software whose settings come from a national institute, InEK, which recalculates the weights each year.
europepmc.orgyalealumnimagazine.compmc.ncbi.nlm.nih.govepub.ub.uni-muenchen.de
Credit
Robert B. Fetter and John D. Thompson, Yale University; US Congress and Medicare
2003
One electronic claim format
HIPAA, a 1996 US law, required national standards for electronic claims and other transactions between providers and insurers; a rule of August 2000 adopted eight, among them the claim known by its number, 837, and every provider and insurer using them had to comply by 16 October 2003 at the latest. A 2001 law also required most Medicare claims to be electronic from that date; in fiscal year 2002 about 139 million, 13.9% of the total, had still arrived on paper. Statistics showed too few providers could send compliant claims by the deadline, so Medicare kept accepting older electronic formats; from 1 July 2004 it paid non-standard claims no sooner than 27 days after receipt, and by 10 September 2004 about 97.7% of incoming claims were in the standard format.
Lessona legal deadline alone did not move providers to the standard; Medicare kept the old formats open and then used slower payment to bring the rest across.
Credit
US Congress; US Department of Health and Human Services and CMS
2018
Paper referrals switched off
Choose and Book was England's national service for booking a first hospital or clinic appointment, with a choice of place, date and time, when a GP (family doctor) refers a patient; the target was that by December 2005 every patient referred for non-emergency hospital care would be offered a choice of four or five providers. In January 2005 the National Audit Office reported that only 60 to 70% of the NHS would have the system by then, that about half of the GPs it surveyed knew very little about choice, and that low GP engagement was a key risk. Its successor, the NHS e-Referral Service (e-RS), still carried only 62% of GP referrals to consultant-led first outpatient appointments in April 2018. NHS England then used its standard contract: after 1 October 2018, hospitals were no longer paid for first outpatient attendances that came from GP referrals sent outside e-RS, and could return such referrals to the practice.
Lessona national booking system that clinicians could bypass stayed partial for more than a decade, and NHS England in the end tied hospital payment to its use.
Credit
Department of Health; NHS England and NHS Digital (Paper Switch-Off Programme)
2022
Overbooking falls on Black patients
Clinics use machine learning to pick out the patients most likely to miss appointments (no-shows) and book them into or right after overbooked slots, which keeps doctors busy but leaves those patients waiting longer. Working from a large specialty clinic whose Black patients had a higher no-show probability, Michele Samorani and colleagues showed that state-of-the-art scheduling made Black patients in their data wait about 30% longer than non-Black patients. A "race-aware" objective, which minimizes the waiting time of the group expected to wait longest, removed the gap at a schedule cost similar to the standard method; methods that left race out did not achieve both.
Lessona schedule tuned only for the clinic's efficiency can load the cost of predicted no-shows onto one group of patients.
Credit
Michele Samorani, Haibing Lu and Michael Santoro (Santa Clara University); Shannon Harris (Virginia Commonwealth University); Linda Goler Blount (Black Women's Health Imperative)
2024
Ransomware stops a claims hub
Change Healthcare, part of UnitedHealth Group's Optum, is a clearinghouse, a company that passes claims and payments between doctors, pharmacists, insurers and governments. On 21 February 2024 a group calling itself ALPHV or BlackCat set off ransomware inside its systems; by the chief executive's own testimony, the intruders had got in nine days earlier with stolen credentials through a remote-access portal that had no multi-factor authentication. In an informal American Medical Association survey from 26 March to 3 April 2024, with more than 1,400 responses, 80% of practices answering said they had lost revenue from unpaid claims, 32% could not submit claims, 55% had used personal funds to cover practice expenses and 48% had signed up with another clearinghouse. Change Healthcare later put the number of people whose data was exposed at about 192.7 million, the figure on the US health department's breach portal.
Lessonwhen one clearinghouse carries a large share of a country's claims, practices with no second connection stop being paid when it stops.
Credit
Change Healthcare and UnitedHealth Group (the company attacked); American Medical Association (survey)
2024
Denials rise with automation
Post-acute care is care after a hospital stay, such as rehabilitation or a skilled nursing facility. In a report of 17 October 2024 built on more than 280,000 pages of documents from UnitedHealthcare, Humana and CVS, which together cover nearly 60% of people in Medicare Advantage (the private plans within Medicare), the US Senate's Permanent Subcommittee on Investigations found that UnitedHealthcare's denial rate for post-acute prior authorization requests rose from 10.9% in 2020 to 22.7% in 2022 while it brought in several automation projects; minutes on one model tested in 2021 recorded "faster handle times" and "an increase in adverse determination rate." In 2022 Humana's post-acute denial rate was more than 16 times its overall rate. The subcommittee asked CMS to collect prior authorization data by type of service, audit insurers whose denial rates jump, and keep predictive tools from unduly influencing human reviewers.
Lessona tool that speeds up review can also raise denials, so its effect shows in denial rates by type of care, not in handling times.
Credit
Senator Richard Blumenthal and the subcommittee's majority staff
Watch for
AI prior approval in Medicare
From 1 January 2026 to 2031, CMS's WISeR (Wasteful and Inappropriate Service Reduction) model has technology companies run prior authorization in traditional Medicare in Arizona, New Jersey, Ohio, Oklahoma, Texas and Washington for selected services, such as skin substitutes, electrical nerve stimulator implants and knee arthroscopy for knee osteoarthritis, using "Artificial Intelligence (AI) and Machine Learning (ML), along with human clinical review." The companies are paid a percentage of the spending on care they avert, and licensed clinicians make every recommendation for non-payment. An outside study in JAMA by Hannah Neprash and Michael Barnett, using Epic Cosmos records, found that in the first six months use of epidural steroid injections fell about 15% and other targeted services about 20% against trends in other states; the authors say they cannot yet tell whether the care not received "was wasteful or necessary."
Achieved whenclinicians independent of CMS and the companies review a sample of WISeR decisions and report how often needed care was refused or delayed.
Claimed by
CMS, which announced the model in June 2025 as a way to cut wasteful and inappropriate care
Insurers pledge simpler prior approvals
On 23 June 2025 dozens of US health plans, among them Aetna, Kaiser Permanente, UnitedHealthcare and the Blue Cross Blue Shield Association, pledged through AHIP, the insurers' trade group, to show reductions in the services needing prior authorization by 1 January 2026, honor existing approvals for 90 days when a patient changes plans, answer at least 80% of electronic approvals in real time by 2027, and have a standard electronic framework working by 1 January 2027; AHIP said the pledge would benefit 257 million Americans. In June 2026 AHIP reported that prior authorization volume had fallen 11% over the past year, about 6.5 million fewer authorizations.
Achieved whenthe prior authorization figures that CMS's January 2024 rule makes Medicare Advantage, Medicaid, CHIP and exchange plans post on their websites each year, or an independent study of claims data, show the fall and the real-time answer rate plan by plan.
Claimed by
AHIP and its member plans, June 2025 and June 2026
AI coding that cuts denials
Epic Systems, which sells electronic health records, says that at the health systems using its billing and coding AI, Penny, most actively, coding-related denials (claims refused because of coding errors) have dropped by more than 20%, denial appeal letters are written 23% faster, and at Summit Health medication prior authorization submission time has fallen 42%.
Achieved whenan outside study using claims data compares denial rates at Penny users with matched non-users, or with the same systems before they adopted it.
Claimed by
Epic Systems, March 2026
Open
01
A prior authorization answer key
A public set of at least 300 synthetic prior authorization requests. Each is a made-up patient from Synthea in FHIR format with a short clinic note written by a clinician, asks for one service, and is tied to one of Medicare's published coverage rules (National or Local Coverage Determinations, NCDs and LCDs, from CMS's Medicare Coverage Database), starting with services on the WISeR list. At least three reviewers with utilization review experience label each case approve, deny or need more information, citing the sentence of the rule they relied on. Then run AI agents and prior authorization products on the same cases, naming them where makers agree, and count wrong denials and wrong approvals separately. No real patients or real requests are used.
The briefCredit · unclaimed02
AI coders tested by name
A public set of at least 200 synthetic outpatient visit notes, each coded independently by two certified medical coders with ICD-10-CM diagnosis codes and the visit's evaluation and management level, with a third coder settling disagreements. Run commercial AI coding products, by name, and open models on the same notes, and report exact matches, near misses (right code family, wrong code), missed codes and added codes. Notes are synthetic; no patient records are used.
The briefCredit · unclaimed03
Coverage rules as code
An open pipeline and dataset that turn Medicare's written coverage rules (LCDs and NCDs) into Documentation Templates and Rules (DTR) packages. DTR is the HL7 Da Vinci format in which an insurer states the documentation it needs as a FHIR Questionnaire, with logic in Clinical Quality Language (CQL) that fills in answers from the patient's record. Each generated rule links to the sentence it came from, and sentences that cannot be encoded are listed with a reason. Packages are tested on synthetic patients only.
The briefCredit · unclaimed
Open 01 · the brief
A prior authorization answer key
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public set of at least 300 synthetic prior authorization requests. Each is a made-up patient from Synthea in FHIR format with a short clinic note written by a clinician, asks for one service, and is tied to one of Medicare's published coverage rules (National or Local Coverage Determinations, NCDs and LCDs, from CMS's Medicare Coverage Database), starting with services on the WISeR list. At least three reviewers with utilization review experience label each case approve, deny or need more information, citing the sentence of the rule they relied on. Then run AI agents and prior authorization products on the same cases, naming them where makers agree, and count wrong denials and wrong approvals separately. No real patients or real requests are used.
Why it is still open
The nearest work, CHI-Bench (May 2026), gives AI agents 25 provider-side and 25 payer-side prior authorization tasks against a managed-care handbook developed with Johns Hopkins Medicine; its cases were written by Claude Opus 4.7 and audited by clinicians, its reference outcomes come from annotators checked by at least one practicing healthcare worker, and the best agent resolved 28.0% of its tasks. Its authors call the task set "relatively small" and leave a strong human baseline to future work, and the handbook is not Medicare's published rules. A March 2026 study scored AI-written prior authorization letters on 45 synthetic scenarios, which tests the letter, not the decision. The first outside study of WISeR could not say whether the care it cut was needed.
Built with what exists
Synthea, an Apache-2.0 synthetic patient generator that writes FHIR R4; the Medicare Coverage Database downloads, CSV files of all NCDs and LCDs updated weekly; CHI-Bench's Apache-2.0 code for running agents through simulated apps; a model built for fixed-answer questions, such as Google Gemini Decision 1, among the systems tested (see the Note "A thousand questions, one pass"); reviewers paid by the hour. The CMS download carries the AMA's license for CPT codes (personal, non-commercial use only), so publish each service by plain description and coverage rule number, not CPT text. Agents that can work through many tools in one task are what make this test worth running now.
First step
Choose three WISeR services that have a Local Coverage Determination, write 30 cases from Synthea patients, and have three reviewers label them to measure agreement before writing more.
Done looks like
At least 300 cases across at least 10 coverage rules, each labeled by 3 reviewers with agreement (Fleiss' kappa) reported; at least 5 AI systems scored on every case, with wrong-denial, wrong-approval and need-more-information rates; cases, labels, scores and code public under an open license.
Open 02 · the brief
AI coders tested by name
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public set of at least 200 synthetic outpatient visit notes, each coded independently by two certified medical coders with ICD-10-CM diagnosis codes and the visit's evaluation and management level, with a third coder settling disagreements. Run commercial AI coding products, by name, and open models on the same notes, and report exact matches, near misses (right code family, wrong code), missed codes and added codes. Notes are synthetic; no patient records are used.
Why it is still open
The US Government Accountability Office's July 2026 spotlight on AI for medical notes and coding says "There are few independent studies evaluating the accuracy of these tools," and the 95% accuracy figure it reports comes from one developer describing its own deployment at a five-hospital system. Research sets do not test products: Soroush and colleagues at Mount Sinai (NEJM AI, April 2024) asked models only to turn code descriptions back into codes, and GPT-4 matched 33.9% of ICD-10-CM codes exactly; Yuan and colleagues (Cambridge and Amazon, EMNLP 2025) released 207 outpatient notes built on ACI-Bench with ICD-10-CM codes only, no procedure codes, annotators' credentials unstated, under a non-commercial license. The larger MIMIC hospital notes may not be sent to outside services that keep the data, under PhysioNet's rules. Makers publish their own figures: Commure said in October 2026 that its AI "coded over 85% of all charges" at one OB/GYN group, with no accuracy figure.
gao.goveurekalert.orgarxiv.orggithub.comphysionet.orgcommure.com
Built with what exists
ACI-Bench visit dialogues and notes, openly licensed (CC BY), as the base; Yuan and colleagues' annotations for comparison; Synthea for extra cases; CMS's ICD-10-CM code files for fiscal year 2027; an open-weight model run locally, such as Mistral Large 4, as a baseline (see the Note "A trillion open parameters"); test access from makers. Visit levels and other procedure codes are CPT codes licensed by the AMA, so publish those results only as scores unless an AMA license is obtained.
First step
Merge Yuan and colleagues' annotations with ACI-Bench, have two certified coders recode 30 notes on their own and add visit levels, compare with the published labels, and ask five makers for test access.
Done looks like
At least 200 notes, each coded by 2 certified coders with their agreement reported; at least 5 named products and 2 open models scored on every note, with per-code precision and recall and the share of near misses; notes and ICD-10-CM answers public under CC BY.
Open 03 · the brief
Coverage rules as code
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
An open pipeline and dataset that turn Medicare's written coverage rules (LCDs and NCDs) into Documentation Templates and Rules (DTR) packages. DTR is the HL7 Da Vinci format in which an insurer states the documentation it needs as a FHIR Questionnaire, with logic in Clinical Quality Language (CQL) that fills in answers from the patient's record. Each generated rule links to the sentence it came from, and sentences that cannot be encoded are listed with a reason. Packages are tested on synthetic patients only.
Why it is still open
The public Da Vinci library holds hand-built rule packages for 10 topics, such as home oxygen therapy, positive airway pressure devices and hospital beds. GuidelineForge, a NeurIPS 2026 workshop paper, turns open clinical guidelines into CQL with every rule tied to its source sentence, but it works on guidelines, not coverage rules, and its page states no code release. Medicare's coverage rules are published as text, in CSV files with embedded HTML. Meanwhile CMS's January 2024 rule requires Medicare Advantage, Medicaid, CHIP and exchange plans to offer a Prior Authorization API from 1 January 2027.
Built with what exists
the Medicare Coverage Database downloads; the Da Vinci DTR guide and the CDS-Library packages as an answer key; the CQL-to-ELM translator from the cqframework project (Apache-2.0) to compile the logic; Synthea test patients; the Inferno Da Vinci CRD test kit (Apache-2.0); a model suited to long-document parsing and extraction, such as Meta Muse Ember, to draft rules for a person to check (see the Note "The jobs agents repeat").
First step
Download the current coverage data, find the Medicare rule on home oxygen, generate a package from its text, and compare its answers with the hand-built HomeOxygenTherapy package on 50 Synthea patients.
Done looks like
At least 50 coverage rules converted; every package compiles; for the 10 topics in the Da Vinci library, the generated packages give the same answer as the hand-built ones for at least 95% of 50 synthetic patients per topic, half of them built to meet the rule; every rule cites its source sentence; packages, refusal lists and code public under Apache-2.0.