AI for healthcare providers: from triage to record AI on your EHR

Hospitals, mental health institutions, long-term care organisations and primary care providers are under pressure: rising demand for care, a tight labour market and growing administrative burden. Artificial intelligence can relieve care professionals, from automatic documentation via speech recognition to predictive models for readmission and no-shows. We build AI applications that comply with NEN 7510, GDPR Article 9 and the classification requirements under the MDR for software as a medical device.

Clinical decision support Record AI on EHR Speech-to-report Triage and planning AI NEN 7510 and MDR

Discuss your healthcare AI case View applications

Why AI in healthcare is more than a buzzword

The healthcare sector has been collecting structured and unstructured data for decades: EHR notes, images, lab values, monitoring signals, claims data. Only with the current generation of language models and machine learning frameworks does this data become practically usable for clinical support, planning and quality improvement, provided the implementation is done carefully.

A typical nurse spends a considerable part of each shift on documentation. An emergency department doctor has to triage within minutes with incomplete information. A mental health clinician works with records where the relevant context is spread across years of separate notes. AI helps in exactly these places: not by replacing the care professional, but by reducing the cognitive load and surfacing patterns that get lost in a manual process. We build AI solutions that deliver measurable time savings and fit existing workflows, including through integrations with custom healthcare software and electronic patient record applications.

Dutch healthcare operates within a complex regulatory framework: GDPR Article 9 for special category personal data, the Healthcare Quality, Complaints and Disputes Act (Wkkgz), the Medical Treatment Agreement Act (WGBO), the Electronic Data Exchange in Healthcare Act (Wegiz) and the Medical Device Regulation (MDR) for software that qualifies as a medical device. In addition, the European AI Act classifies many healthcare applications as high-risk. A workable AI solution in healthcare takes these frameworks into account from the architecture stage onwards, not as a compliance layer added afterwards.

Core areas where AI makes a difference for healthcare organisations

We work on use cases that save time, raise quality and keep costs under control. For each area, we build custom solutions that fit your EHR, workflow and compliance framework.

Clinical decision support

Models that alert clinicians to abnormal lab values, medication contra-indications or patterns in monitoring data. Not autonomous diagnosis, but a second pair of eyes that watches around the clock and generates alerts within the care provider's workflow.

Record AI on EHR and ECD data

Summaries, automated problem lists and search across years of unstructured notes. Works with HiX, ChipSoft, Epic, Nexus, Pluriform and open-source EHRs via HL7 FHIR integrations, using retrieval-augmented generation under strict data isolation.

Readmission and decompensation prediction models

Machine learning models that use historical record data to estimate risk: which patient is likely to be readmitted within 30 days, which COPD patient is on an upward decompensation curve, which mental health client is showing signs of crisis.

Triage AI in the emergency department

Decision support for initial assessment that combines symptoms, vital signs and historical data into a priority indication. Complements established triage systems, with full logging under NEN 7513 for retrospective traceability.

Image recognition in radiology and dermatology

Computer vision for screening X-ray, CT and MRI images or dermatological photos. We don't build models from scratch when certified alternatives exist — we integrate existing tooling and develop custom solutions where the market offers no suitable option.

Speech-to-report for care providers

Whisper-based transcription that turns patient conversations straight into a structured report, filed in the right place in the EHR. Can be deployed fully on-premise or in a Dutch cloud environment, with dynamic redaction of direct identifiers.

Specific applications by type of care organisation

A hospital faces different challenges from a mental health institution or a GP network. We tackle the use cases that will deliver the most value in your specific context.

Hospitals and university medical centres

Capacity planning for operating theatres and outpatient clinics, forecasting bed occupancy by department, automatic coding of care activities to DBC/DOT, reminder flows for invoices, fraud detection on claims streams, and AI support in radiology workflows. Integrations with HiX, Epic and ChipSoft via HL7 FHIR and local Twiin nodes.

Mental health institutions

Sentiment and risk analysis of treatment notes (with explicit clinician confirmation), automatic generation of treatment plans from medical history, no-show prediction for outpatient appointments and intelligent waiting list prioritisation. Our mental health software and the AI components described here can be combined in a modular way.

Residential and home-care organisations

Prediction models for fall incidents and decompensation in clients with chronic conditions, smart scheduling for community nursing based on care intensity and travel time, speech-to-report for carers on the move, and sentiment monitoring of client feedback to detect quality issues early.

Primary care and care cooperatives

AI intake for GP out-of-hours services with automated triage, no-show prediction for consultations, automatic consultation notes via speech recognition, and integrations with GP systems such as Medicom, Promedico and CGM. For healthcare groups we also offer BI dashboards and predictive analytics on population data.

Compliance is the foundation, not the finish line

In healthcare, an AI application without a solid compliance foundation is not production-ready. We design AI systems in which privacy, security and MDR classification are part of the architecture from day one, rather than a checklist ticked off afterwards.

NEN 7510, 7512 and 7513

NEN 7510 for information security in healthcare, NEN 7512 for secure authentication of care providers (with a UZI card or equivalent) and NEN 7513 for logging and audit trails of access to patient data. Our infrastructure and application architecture are built to these standards.

GDPR Article 9 and BSN policy

Health data are special category personal data under Article 9 of the GDPR. Processing requires explicit legal bases, purpose limitation and strict data minimisation. The BSN may only be processed within the framework of the Wabb and the Wet aanvullende bepalingen verwerking persoonsgegevens in de zorg.

Wegiz, MedMij and VIPP

The Electronic Health Data Exchange in Healthcare Act (Wet elektronische gegevensuitwisseling in de zorg) requires standardised exchange between healthcare providers. We build integrations that comply with the MedMij framework, VIPP requirements and the LSP/AORTA framework, with FHIR R4 as the baseline standard.

MDR and the AI Act for SaMD

Software that qualifies as a medical device falls under the Medical Device Regulation (MDR) and must be assessed by a notified body. The AI Act classifies many healthcare applications as high-risk. We advise on classification early on and help prevent a prototype from unintentionally falling within MDR scope.

More on AI in healthcare Healthcare app development Building a healthcare app

How we build AI for healthcare organisations

An AI project in healthcare succeeds or fails on data quality, buy-in from care professionals and compliance discipline. Our approach builds in each of these three explicitly.

We typically begin with an AI discovery workshop in which we work with care professionals, IT and the privacy officer to identify the most promising use cases. Not the most spectacular, but the most feasible with demonstrable value. We then build a proof of concept on an anonymised or synthetic dataset to validate the assumption before investing time and money in a production track.

During the build phase we work in short sprints with direct feedback loops from end users. An AI tool that is technically perfect but not used by nurses has no value. We deliberately integrate deeply with the existing EHR via FHIR integrations, single sign-on and the healthcare provider's standard notification channel, rather than a separate application alongside the workflow. For specific domains we work with clinical experts and can provide an interim AI tech lead who also brings the organisation along internally.

From first conversation to production-ready AI

A structured programme of four to six months, depending on complexity and MDR classification, with clear go/no-go decision points.

Discovery and use case prioritisation

A workshop with care professionals, IT, the privacy officer and, where relevant, the medical ethics committee. We map opportunities, risks and MDR classification, and prioritise by feasibility and impact.

Data assessment and privacy design

We inventory available data sources, assess quality and completeness, and design the privacy framework: pseudonymisation, purpose limitation, retention periods and logging in accordance with NEN 7513.

Proof of concept on controlled data

We build a working prototype on synthetic or anonymised data. Here we validate technical feasibility and the business case before entering the production track.

Production deployment and integration

Rollout within your EHR environment via FHIR integrations, integration with SSO and authorisation structures, monitoring, audit trails and end-user training, with explicit go-live criteria.

Technologies and standards we use

We combine current AI frameworks with the specific standards and interfaces of Dutch healthcare.

Healthcare interoperability

HL7 FHIR R4 as the baseline standard for exchange. Integrations with Twiin nodes, the LSP/AORTA framework, MedMij data services and VIPP flows. DICOM for imaging, and EDIFACT and HL7v2 for administrative exchange where still applicable.

HL7 FHIR R4 DICOM Twiin LSP / AORTA MedMij VIPP

EHR and EPD integrations

Hands-on experience with integrations towards ChipSoft HiX, Epic, Nexus, Pluriform Medical, Medicom, Promedico, CGM, Ysis and open-source EHRs. We work through certified interfaces, respecting the vendor's software lifecycle management.

HiX / ChipSoft Epic Nexus Medicom Promedico Ysis

AI and ML stack

For language models we work with OpenAI, Anthropic Claude, and open-source alternatives such as Llama and Mistral where on-premise deployment is required. For classical ML we use scikit-learn, XGBoost, PyTorch and TensorFlow. Whisper for speech-to-text; for healthcare Dutch we fine-tune where necessary.

OpenAI / Anthropic Llama / Mistral Whisper PyTorch scikit-learn

Hosting, security and logging

Deployment in Dutch cloud environments, sovereign private clouds or on-premise, depending on the sensitivity of the data. ISO 27001-compliant infrastructure, processes aligned with NEN 7510, audit logging under NEN 7513, and authentication with UZI cards or comparable PKIO means.

NEN 7510 NEN 7512 NEN 7513 ISO 27001 UZI / PKIO

Why choose Appfront for AI in healthcare

Healthcare calls for a partner who understands both the technical side and the regulatory and organisational context. We combine hands-on AI engineering with a pragmatic view of feasibility.

Honest about what AI can and cannot do

We promise no miracles. A hallucinating language model has no place in a clinical workflow without human oversight and clear fail-safes. We identify risks explicitly and build in redundancy where needed.

Broad experience beyond healthcare

Our experience in financial services, government and logistics yields patterns that are valuable in healthcare: risk modelling, document AI and planning optimisation are cross-sector disciplines. See also our AI for banks and AI for municipalities.

End-to-end responsibility

From use-case discovery to production deployment, with the same teams. We build not only the AI model but also the integrations, the user interface and the observability layer. We do not hand over ultimate responsibility to your IT department.

A pragmatic Dutch mid-sized firm

Not enterprise rates for a simple form, but also not freelance flexibility without governance. We work transparently, in sprints with fixed moments for demo and review, and stop in good time when a use case does not work in practice.

Concrete scenarios we work with

No invented client cases, but patterns we encounter often and which illustrate what is technically and organisationally feasible.

Speech-to-EHR for nurses

A mobile app on a care tablet that transcribes conversations with clients, structures them and places them in the right section of the EHR. Whisper runs in a sovereign cloud, redaction happens before storage, and the nurse always confirms the report before it is saved.

No-show prediction for outpatient clinics

An ML model that estimates a no-show probability per appointment based on historical data: age, previous no-show behaviour, appointment type, weekday, weather and travel distance. High-risk appointments receive extra reminders or a phone-call workflow from the clinic assistant.

Readmission risk at discharge

A dashboard that, during discharge planning, shows a 30-day readmission probability for each patient, with the five main drivers per patient. The doctor decides — the model advises, with an explicit uncertainty indicator and an audit log under NEN 7513.

Claims validation and fraud detection

Anomaly detection on claims streams, focused on errors and inconsistencies — not on casting suspicion on care providers. The model flags deviating patterns for manual review by the claims administration.

AI intake for out-of-hours GP service

A conversational AI that makes an initial sort based on complaints and suggests the right questions for the triage nurse, connected to a GP system. No autonomous triage, but a faster and more consistent intake process. See our AI intake integration.

Sentiment analysis of patient feedback

Automated analysis of free-text fields in patient satisfaction surveys, with thematic clustering. Your quality department can see trends by location and theme, faster than manual coding, with traceable added value.

Start small, validate rigorously, then scale

The biggest mistake in healthcare AI projects is starting too big, with too broad a scope and too many stakeholders. Our approach: choose a use case where the value is measurable within three months, validate with real users on real data, and only then invest in a full rollout. A working AI feature for a specific workflow is worth more than an ambitious platform that never reaches production. We help you weigh this up honestly, and we'll tell you "you shouldn't do this now" when that is the right advice.

Frequently asked questions about AI for healthcare organisations

Does our AI application fall under the MDR as a medical device?

That depends on the intended purpose. Software that supports diagnostic or therapeutic decisions often falls under MDR classification and requires assessment by a notified body. We map the classification early and, where possible, help you design the scope so that a use case deliberately sits outside the MDR. If MDR classification is unavoidable, we guide you through the process.

How do you handle patient data when training models?

For proofs of concept, we work by default with synthetic or anonymised data. In production, we apply strict pseudonymisation, with logging in line with NEN 7513 and clear purpose limitation under GDPR Article 9. Models are not trained on personal data without an explicit legal basis, and we never use healthcare data to improve public models.

Can this be integrated with our EHR, such as HiX or Epic?

Yes. We work with the certified interfaces of the leading EHR vendors (Chipsoft HiX, Epic, Nexus, Pluriform), using HL7 FHIR R4 as the baseline standard. We also integrate with GP systems such as Medicom, Promedico and CGM, and with long-term care systems such as Ysis. For each project, we agree the integration plan with your EHR vendor.

Do you work on-premise or in the cloud?

Both. For applications involving highly sensitive data or strict data residency requirements, we work on-premise or within a sovereign private cloud in the Netherlands. For less sensitive use cases, we use a cloud environment aligned with NEN 7510. We decide case by case, based on data classification, MDR status and your own IT architecture.

How do we prevent a language model from hallucinating in a clinical context?

Through retrieval-augmented generation: the model only has access to source data from the patient record and must provide citations back to the source. We also build in explicit fail-safes, so the model declines to give an answer when uncertain, and human confirmation is always required before information is entered into the record. AI does not replace clinical responsibility.

What determines the investment in an AI project for our healthcare organisation?

The biggest factors: the complexity of the use case, the data quality, the number of integrations with existing systems, and whether the project falls within the scope of the MDR. A speech-to-report prototype is a fundamentally different project from a production-ready readmission predictor with MDR certification. We always start with a focused discovery workshop before preparing a tailored quote.

How long before we see results?

A proof of concept typically takes six to ten weeks. Production implementation depends heavily on the complexity of EHR integrations and compliance requirements. Four to six months is realistic for a use case outside MDR scope. We work in iterative sprints with interim demos, so you can steer the project as it progresses.

Do you build it yourselves or supply ready-made products?

We build custom. Where certified standard products exist for a use case (for example, radiological image recognition), we advise integrating them rather than building our own. Our added value lies in the combination: custom AI on your data and work processes, combined with proven building blocks where that makes sense.

Exploring AI opportunities for your healthcare organisation?

Schedule a no-obligation conversation with our healthcare AI specialists. We'll discuss your situation, map out promising use cases and be honest about what is (not yet) feasible.

Book a no-obligation conversation