AI-ondersteund coderen
AI-ondersteund coderen is the Dutch Hospital Data AI model that reads clinical texts in a hospital's electronic patient record and assigns the principal ICD-10 diagnosis to day admissions, freeing medical coders in 32 Dutch hospitals from repetitive work.
What is AI-ondersteund coderen?
AI-ondersteund coderen, shortened to AIOC by its publisher, is the artificial intelligence model built by Stichting Dutch Hospital Data (DHD) together with a collective of twenty hospitals. Its job is narrow and concrete: reading the clinical text already held in a hospital's electronic patient record, it assigns the principal ICD-10 diagnosis to day admissions, the dagopnamen that make up a large share of the roughly three million admissions coded in the Netherlands every year. Coding them is a statutory chore, and qualified medical coders are in short supply. The model rests on natural language processing algorithms trained on large volumes of medical data. It reads specialists' letters and reports, recognises the patterns that point to a diagnosis, and attaches the matching code together with a certainty percentage. Training drew on more than 300,000 day admissions and over a million source documents, contributed by four training hospitals across five years of registration. What sets AIOC apart is how that training happens. Learning is federated: the model travels to servers inside each hospital rather than the records travelling out to a central store. The underlying technology is vantage6, software from IKNL, combining a central server at DHD, local nodes and Docker containers. The product sits inside PLUGIN, the Platform voor Uitwisseling en Hergebruik van Klinische Data Nederland, an initiative of DHD, the Expertisecentrum Zorgalgoritmen and IKNL. AIOC has been in production since the start of 2024. Thirty-two hospitals now use it and 165,000 day admissions were coded with it during that year. Statistics Netherlands, the CBS, has judged the output fit for its statistical applications and accepts codes above 70% certainty without a human check; measured accuracy sits between 85% and 90%, comparable to that of a medical coder. On average the model codes around 65% of day admissions with very high certainty, leaving the remainder to people. A companion model for clinical admissions was announced to start in 2025.
What it does
- Assign the principal ICD-10 diagnosis to day admissions automatically, straight from the clinical text
- Return a certainty percentage for every code, so a hospital can see what may be accepted without a manual check
- Take repetitive coding off the desks of medical coders and free them for clinical admissions
- Raise the quality, consistency and completeness of diagnosis coding
- Feed the national LBZ hospital care registry through the ordinary delivery route
- Write the coding results back into the electronic patient record via Epic or ChipSoft HiX
- Improve itself on hospital data locally, without those records ever being centralised
When to use AI-ondersteund coderen / When not to
A quick filter to help you decide if AI-ondersteund coderen is the right fit.
When to use AI-ondersteund coderen
- Dutch hospitals that have to code the diagnoses of their day admissions for the national CBS statistical obligation
- Health record administration departments (zorgadministratie) squeezed by the shortage of qualified medical coders
- Hospitals running Epic or ChipSoft HiX 6.3 and later, where the link back into the electronic patient record has already been built
- Hospital IT departments able to run and maintain a dedicated Linux server with Docker inside their own network
- Medical coders who would rather spend their time on complex clinical admissions than on repetitive day admissions
When not to use AI-ondersteund coderen
- Care organisations outside the Netherlands: the model is currently offered to Dutch hospitals only
- Teams that need clinical inpatient admissions coded, since the model covers day admissions alone
- Coders who need secondary diagnoses as well, because only the principal diagnosis is assigned
- IT departments confined to Windows or without Linux and Docker skills, as DHD offers no help with installing or managing Linux servers
- Individual clinicians and non-hospital settings, since access runs through a signed institutional agreement and a local administrator
How to use AI-ondersteund coderen
A typical end-to-end flow, from setup to results.
- Contact DHD at ai-coderen@dhd.nl to discuss the fit, the current rates and the agreements to be signed
- Decide the role your hospital will play: an inference hospital that only uses the model, or a training hospital that also trains it and carries heavier hardware requirements
- Provide a local server, a virtual machine being acceptable, running Ubuntu 22.04 or later with Docker available and access to the data; Linux is mandatory and Windows is not possible
- Install a vantage6 node on that server and set up the prescribed folder structure, following the seven-chapter installation guide published online
- Whitelist the DHD IP addresses and create the personal accounts DHD needs for occasional supervision
- Extract a set of day admissions from the electronic patient record onto the local server and pseudonymise it
- Let DHD send the model to that server; results are written back to the same machine
- Load the results into the electronic patient record through the Epic module or the ChipSoft link for HiX 6.3 and later
- Have the medical coders check, correct and complete the codes, using the certainty percentage to prioritise their review
- Deliver the finished coding to the national LBZ registry through the usual route, and manage user access through your local administrator in the Mijn DHD portal
Pros & Cons
Pros
- Measured accuracy of 85% to 90%, comparable to a medical coder, and endorsed by Statistics Netherlands for codes above 70% certainty
- Around 65% of day admissions coded with very high certainty; Rijnstate reports 72% automatic coding
- Patient data never leave the hospital, thanks to federated learning and pseudonymisation at source
- Real and measurable adoption: 32 hospitals and 165,000 day admissions coded in 2024, in production since early that year
- Electronic patient record integrations already built by Epic and ChipSoft, so no bespoke interface project
- Published by a not-for-profit foundation owned by the sector itself, through the NVZ, UMCNL and the Federatie Medisch Specialisten
- ISO/IEC 27001 and NEN 7510 certifications held by the publisher and externally audited, with a validation report available on request
Cons
- No public price at all: an annual contribution with a fixed and a variable component, plus one-off installation costs, quoted only on request
- Restricted to hospitals in the Netherlands
- Confined to day admissions and to the principal diagnosis, leaving clinical admissions and secondary diagnoses out
- Heavy technical prerequisites, with an Ubuntu 22.04 Linux server, Docker and IP whitelisting, and no DHD support for running Linux
- Roughly 35% of day admissions still need to be checked or coded by hand
- Website and documentation in Dutch only, with neither terms of service nor a pricing page published anywhere on the site
- The NEXUS integration is only at discussion stage, and the Trust Center at compliance.dhd.nl cannot be opened without clearing a Cloudflare challenge
Pricing & Plans
There is no free plan and no free trial for this product, and no price is published. For the day admissions model DHD applies an annual contribution made up of a fixed and a variable component, together with one-off installation costs; current rates are supplied on request only, by writing to ai-coderen@dhd.nl. Prospective users should budget for indirect costs as well, since the hospital supplies the Linux server the model runs on and may need an external IT provider where in-house skills are lacking, DHD offering no support for Linux installation or administration. Access to DHD products is granted by the institution's own local administrator.
Data, GDPR & hosting
A consolidated view of how AI-ondersteund coderen handles your data.
GDPR overview
GDPR implementation is described in concrete terms. DHD acts as a processor under Article 4(8) GDPR for hospital data and as controller only for its own website, and signs a data processing agreement (verwerkersovereenkomst) with every hospital. Its external privacy statement, version 1.1 dated 11 September 2025 and published as a public document, states that processing follows applicable law including the GDPR. A data protection officer is named, Klaas-Wouter Geleynse, at fg@dhd.nl, alongside a three-member advisory privacy committee. DHD says it is certified against the current versions of ISO/IEC 27001 and NEN 7510, audited by an external party. ISO/IEC 27701 is only announced as future work, while ISO 27002, NEN 7512 and NTA 7516 are applied without certification. DHD is also a member of Z-CERT.
Who owns the data?
The hospital owns its data throughout. DHD's own installation guide states that the sensitive, person-related information the model works on remains the property of the hospital at all times, and that this data stays under the hospitals' management. In legal terms DHD presents itself as a processor within the meaning of Article 4(8) of the GDPR, acting on the instructions of each hospital, which remains the controller. DHD may therefore use the data only to run and support the coding service under the data processing agreement signed with the institution, and patients exercise their rights with their own hospital rather than with DHD.
Reuse rights
Processing takes place exclusively inside the secure ICT environment of the hospital concerned, so patient data never leave the institution. The hospital pseudonymises its data before DHD can reach anything, and DHD has no view of identifiable patient information. Only the trained model travels back to DHD for further work; the records themselves never do, and the model processes them without permanent storage. The output belongs to the hospital and can be used freely: coders check, correct and complete the codes, then deliver them to the national LBZ registry through the usual route, with no separate permission from DHD. DHD staff obtain occasional access to the local server, with a personal account and password over a connection from whitelisted DHD IP addresses, purely to confirm that processing ran correctly and to troubleshoot.
Data retention & training
Hosting summary
There is no SaaS hosting to describe. Processing runs on a local server belonging to the hospital, inside its own secure ICT environment, so patient data never leave the institution. Only the trained model travels back to the DHD environment; the underlying records do not. DHD operates the central vantage6 server, which handles authentication, authorisation and communication with the local nodes, and that traffic runs over an encrypted https REST API and websockets. DHD's own access to a hospital server is restricted to whitelisted IP addresses with a personal account and password, and is used only to verify processing and troubleshoot. Neither a hosting country nor a hosting region is published anywhere on the site, and none can be inferred from the fact that the product is offered to Dutch hospitals alone, so none is stated here.
Where AI-ondersteund coderen works
Country-level availability.
Available in
Things to keep in mind
Risks and trade-offs to weigh before adopting AI-ondersteund coderen.
- Automation bias: with roughly 65% of day admissions coded at very high certainty, a team can slide into waving output through and stop giving the remaining cases the scrutiny they still need.
- Deskilling: if junior coders only ever review machine suggestions, the judgement needed for difficult cases, and for the clinical admissions the model does not cover, may quietly erode.
- Accuracy is 85% to 90%, not 100%. Errors flow onward into the LBZ registry and into national statistics, where they become hard to trace back to the admission that produced them.
- Scope confusion: only day admissions and only the principal diagnosis are covered. Assuming broader coverage would leave records quietly incomplete.
- ISO/IEC 27001 and NEN 7510 are claimed as held and externally audited, but no certificate number or certifying body is published, and the Trust Center that would carry the evidence sits behind a challenge page that cannot be cleared.
- ISO/IEC 27701 is only announced as future work, and NEN 7512 and NTA 7516 are applied rather than certified: treating any of the three as an obtained certification would overstate the assurance on offer.
- Details drift out of date. The published adoption figures are from 2024 and the clinical model was announced for 2025; two legal entities share the publisher's address, one of which the site never names; and the Oudlaan 4 address lapses when DHD moves in January 2027.
Setup & Integrations
Technical difficulty
Setup is decidedly technical and falls to the hospital's IT department. It needs a dedicated server, a virtual machine being acceptable, on Ubuntu 22.04 or later with Docker Engine available; Linux is mandatory and Windows impossible, and Docker Desktop is ruled out as chargeable for organisations and unsuited to production. Extracting and pseudonymising data from the electronic patient record, whitelisting DHD's IP addresses and preparing accounts and folders all sit with the hospital, and a training hospital faces heavier demands than an inference hospital. DHD supports neither the installation nor the administration of Linux servers.
Deployment
Integrations
Supported languages
Behind AI-ondersteund coderen
Social
Resources
All the official URLs gathered for verification and reference.
Frequently asked questions
What is AI-ondersteund coderen?
How accurate is the model?
What share of day admissions is coded automatically?
How does the model actually work?
What does federated training and application mean here?
What is the difference between an inference hospital and a training hospital?
Is patient privacy preserved?
Do the results go back into the electronic patient record?
What does it cost?
What if our hospital has no Linux expertise, and can any organisation join?
Should you pick AI-ondersteund coderen?
AI-ondersteund coderen is a narrow tool that is unusually mature within its narrow scope. It has been in production since the start of 2024, runs in 32 Dutch hospitals, coded 165,000 day admissions in a single year, and carries something few AI products in healthcare can show: the explicit blessing of a national statistics institute, which accepts its output above 70% certainty without a human check. Measured accuracy of 85% to 90% puts it on a par with the coders whose repetitive workload it is meant to absorb. Its governance is equally unusual. The publisher is a not-for-profit foundation owned by the hospital sector itself, which changes the incentive structure around a product handling sensitive clinical text. The federated architecture meets the confidentiality objection head-on rather than arguing around it: the model goes to the data, the data never go to the model, and only the trained model returns to DHD. The reservations are just as clear. The geographic and functional perimeter is tight, covering Dutch hospitals, day admissions and the principal diagnosis alone. The technical prerequisites are real: a dedicated Linux server, Docker, IP whitelisting, data extraction and pseudonymisation, none of which DHD will run for you. And no price is published anywhere, so the annual contribution and installation costs have to be prised out by email before any comparison is possible. For a Dutch hospital already on Epic or ChipSoft HiX, with an IT department comfortable with Linux and a coding backlog it cannot staff, this is a serious and well-documented option. For anyone else, it is simply not on offer.
- Choosing a selection results in a full page refresh.
- Opens in a new window.