Skip to content
DataDream
← All articles
Healthcare7 min

AI in healthcare: not a replacement but support

Laurens van Dijk, oprichter van DataDream

Laurens van Dijk

Agentic Engineer, DataDream

Share

At conferences the talk is about AI that spots cancer before a radiologist does. In daily practice the first gains sit somewhere else: in the report after a consultation, the coding of a treatment and the question of which patient to call back first. Less spectacular, but that is where time is lost today that could have gone to the patient.

In this piece I go through what AI already does in healthcare, where its limits are, and how a care organisation can start sensibly.

What does AI mean for healthcare professionals?

AI mainly changes the administrative part of the work: reporting, coding, triage and call-back arrangements. That is where doctors and nurses lose time, and where AI can help without putting a patient directly at risk. AI that assists with a diagnosis is a different matter, with much stricter validation requirements.

The technology itself is usually not the problem. The hard part is embedding it: the link with the electronic health record (HiX, Epic or Nedap), processing that complies with the GDPR, and clinicians who actually want to use the tool.

Explainability

One requirement is firm: you must be able to see how an AI reached its conclusion. A system that reports "elevated risk" without explanation does not help a doctor and only makes the question of liability harder. For administrative AI the bar is lower; for diagnostic AI explainability is a precondition.

This is a national topic too. In the AI4HEALTH programme, led by Prof. Carl Moons of UMC Utrecht, dozens of partners work on applying and scaling AI in healthcare, and on guidelines for its responsible use.

AI as a tool

AI does not take over a doctor's work. Healthcare needs someone who carries responsibility, knows the patient's context and can stop when in doubt. That is part of the profession.

AI also makes mistakes, and sometimes different mistakes from a person. That is exactly why the doctor stays in charge and checks what the system proposes.

Cardiologist Eric Topol wrote a book about this with a telling subtitle: Deep Medicine: How Artificial Intelligence Can Make Healthcare Human Again. His point: when AI takes over admin, the clinician gets time back for the patient. In the short term that is the biggest gain.

How is AI already used in healthcare?

Diagnostics and treatment

With the WondGezond app from Radboudumc, patients recovering from surgery take a photo of their wound at home. The app uses AI to check whether the wound is healing well, and with a good result the patient is notified straight away. In radiology, AI systems help assess X-rays and CT scans. None of these applications work without a clinician behind them, and that is how it should be.

In 2020 Nature published a study of an AI system from Google Health that read mammograms for breast cancer. Compared with six radiologists, the system outperformed each of them individually. A study is different from daily practice, but it shows where things can go.

Predicting from health records

Researchers at Mount Sinai in New York built Deep Patient, a model that analysed the electronic health records of about 700,000 patients to predict which diseases would develop. Such a prediction is only useful if you know what to do with it in the work of the ward.

Speed in stroke care

StrokeViewer from Nicolab analyses CT scans of patients with a suspected stroke and helps doctors decide on treatment faster. Here time saved is directly a medical gain, because in a stroke every minute counts.

Where healthcare AI gets stuck

In the Netherlands, AI in healthcare stands or falls on three things: the quality and protection of data (GDPR and the NEN 7510 standard for information security), explainability, and liability for mistakes. The legal side of that last point is not fully settled, which makes boards cautious.

On top of that comes the link with the health record system. A tool that does not write back to HiX or Epic becomes, for most institutions, one more screen nobody logs into. So ask every vendor: does it write back, in which format, and who maintains that link?

What AI cannot do

AI in healthcare is impressive, but not infallible. A model that recognises patterns does not know the full context of a patient: the home situation, earlier complaints, what the partner said on the phone. AI shows no empathy and makes no ethical decisions. More data does not solve that. That is why the healthcare professional stays indispensable.

Where it is heading

A realistic picture is not AI making the diagnosis while the doctor clicks OK. It is closer to this: AI prepares, the doctor decides, AI follows up.

Tailored treatment

AI can help match a treatment more closely to a patient, based on earlier treatments, lab results and literature. That is not equally relevant for every condition. In oncology and chronic care, research is furthest along.

Less admin

When AI takes over administrative tasks, clinicians keep more time for contact with patients. In the short term that is the biggest gain, and it is easily underestimated because it is not spectacular.

Detecting disease earlier

AI can help find disease earlier, sometimes before clear symptoms appear. Whether that helps depends on what happens next: an early signal only matters if there is time and capacity to follow it up.

Where you can start today

Next to the big applications in radiology and research, there are practical uses a care institution can start with now.

Reporting. Tools such as Augmedix, Suki and Nuance's Dragon turn a consultation or a note into a report the doctor checks. Measure beforehand how much time reporting takes now, so that after a trial you know what it delivers.

Triage and planning. A system that classifies incoming complaints and routes them to the right place can help an out-of-hours GP service or emergency department see the right patient first. A clinician keeps the final word.

Patient communication. Answers to questions that keep coming back, about medication, appointments or preparing for a test, through a secure patient portal. Also outside office hours, without someone having to sit there for it.

Coding and registration. A suggested DBC or ICD-10 code based on the record, checked by a staff member. That saves time and makes registrations more complete.

GDPR and NEN 7510

Health data counts as a special category of personal data under the GDPR. On top of that, the NEN 7510 standard for information security in healthcare applies. Cloud AI with patient data needs at least a data processing agreement with the vendor, and patients must know what happens to their data. If the data leaves the EEA, that is only allowed to a country with an adequate level of protection or with extra safeguards, such as standard contractual clauses. For the US it depends on whether the vendor takes part in the Data Privacy Framework (Dutch Data Protection Authority).

Many care organisations add their own procurement requirements on top: storage within the EU and no model training on their data. The GDPR does not require that; it is the organisation's own choice, and a sensible one for patient data. A vendor without a clear answer on these points is not a vendor for healthcare.

For the most sensitive records, such as in youth care or mental health, consider a setup on your own servers or in a closed-off European cloud. Decide that per application.

Start small

AI in healthcare is a tool. In the short term most of the gain sits in the layer below the spectacular applications: admin, coding, triage and patient communication. That is where you can give time back to the people who deliver care today.

Start with a process that visibly costs time, often reporting. Include the health record integration from day one and make the GDPR and NEN 7510 hard requirements when you choose a vendor. Put one tool live on one ward and expand from there.

Curious how AI can support your care organisation? Take the free AI Scan for a first analysis.

Curious what AI can do for your business?

Take the free AI Scan and find out in 1 minute.

Frequently asked questions

Will AI replace the doctor?
No. Healthcare needs someone who carries responsibility, knows the patient's context and can stop when in doubt. AI sometimes makes different mistakes from a person, so the doctor stays in charge and checks what the system proposes. AI prepares, the doctor decides, AI follows up.
How is AI already used in Dutch healthcare?
In diagnostics, such as Radboudumc's WondGezond app, assessment of X-rays and CT scans in radiology and Nicolab's StrokeViewer for strokes, always with a clinician behind it. And in admin: reporting with tools such as Augmedix, Suki and Dragon, triage and planning, patient communication and coding.
Which rules apply to AI in healthcare?
Health data counts as a special category of personal data under the GDPR, and the NEN 7510 standard for information security applies as well. Cloud AI with patient data needs at least a data processing agreement and openness towards the patient. If data leaves the EEA, that is only allowed to a country with an adequate level of protection or with extra safeguards, such as standard contractual clauses. Many care organisations also require EU storage and no training on their data themselves. For the most sensitive records, such as in youth care or mental health, decide per application on a setup on your own servers or in a closed-off European cloud.
Where does AI save the most time in healthcare?
In admin. Reporting is often the best place to start, followed by triage, planning, patient communication and coding. Measure beforehand how much time a process takes now, so that after a trial you know what it delivers, and include the health record integration from day one.