Yesterday I opened MongoDB’s first EMEA-wide healthcare webinar with a question that sounds almost too simple: as healthcare changes, what must our applications know about the patient?

I had twenty minutes before handing over to Francesc Mateu, our Field CTO for Healthcare, and then to a live demo of an intelligent patient journey. Twenty minutes is enough to make one argument, not to prove it. This post is that argument, written down, with the numbers I used and the picture I kept coming back to: a ship, five winds, and a coastline the ship is trying to reach.

The pressure is structural, not cyclical

The starting point is capacity. In 2025, people aged sixty-five and over represented 22% of the European population. Longer lives usually mean more conditions to manage at the same time, more treatments that interact, and more care teams that need to agree on what is going on. Each of those is a coordination problem before it is a clinical one.

Put that next to the second number: by 2030 the world is heading into a shortage of nurses, with demand well above what the training pipeline can deliver. Fewer clinicians will be looking after more complex patients. There is no scenario in which this is solved by asking the people in the system to work harder. Every minute a clinician spends searching for a result, or reconstructing a history from three portals and a phone call, is a minute taken from a shrinking budget of attention. That, more than any technology trend, is why digital health investment keeps rising: the alternative is a care model that quietly stops scaling.

Europe is moving on two levels

The good news is that this is not a greenfield. Europe is already moving, and on two levels at once.

At the Union level, MyHealth@EU is live, exchanging cross-border patient summaries and ePrescriptions between participating countries. The European Health Data Space gives that exchange a legal spine and a calendar: a first phase in 2029 covering patient summaries, prescriptions and dispensations, and a next phase in 2031 extending to medical images, laboratory results and discharge reports. For anyone building healthcare applications in Europe, those two dates are now part of the requirements, whether or not they appear in the tender.

At the country level, the services already run at scale. Germany has redeemed around 1.6 billion e-prescriptions. Finland’s national archive takes in more than two million new documents a day. Estonia issues practically 100% of its prescriptions digitally. Spain has crossed 1.5 million cross-region dispensations. Italy’s Fascicolo Sanitario Elettronico is available to millions of citizens. These are not pilots. They are the plumbing of daily care in their respective countries.

What strikes me, looking at them together, is how similar the forces shaping them are. Different countries, different procurement cultures, different vendors, and yet the same five winds.

Five winds, one sail

During the webinar I used an image that I find more useful than any architecture diagram at this stage of the conversation. Think of the care model as a ship. Five forces are blowing on it. They come from different directions and they have different names, but they fill the same sail and they push the ship towards the same coast.

Interoperability. The patient has to be followable across providers. A summary written in one hospital must mean the same thing in another, in another region, in another country. This is the wind that EHDS turns from aspiration into obligation.

Real-time care. A new laboratory result or a reading from a home device has to reach the right workflow when it matters, not in the next overnight batch. The value of an event decays fast; a critical potassium level at 3 pm is a different fact at 9 am the following day.

Precision medicine. Building a richer picture of the individual means combining clinical information with molecular data: genomics, multi-omics, structured and unstructured, evolving at a completely different pace from an encounter record.

AI. Models and agents only produce useful answers and actions when they are given relevant patient evidence. Without it they are fluent and wrong. I wrote about this at length for agents in general; in healthcare the stakes make the point for me.

Trust. Evidence has to be protected in line with sovereignty requirements, and the system has to stay up. In public healthcare, trust is not a feature on a slide. It is the licence to operate.

Read the five together and a shared requirement emerges that none of them states on its own: an application needs to understand what is happening now in the context of the person’s history. Not a snapshot. Not the latest record. The whole story, with the newest chapter attached.

The coast: longitudinal patient context

The coast the ship is sailing towards has a name in the industry: the longitudinal patient record. I prefer longitudinal patient context, because a record is something you file and a context is something you use. It means following one person across encounters, across providers, and over time, so that each new event adds to a connected history that can support the next decision.

What has to live in that context? More than most people expect. The clinical record itself, encounters and medication. Diagnostics: laboratory results, imaging and reports. Precision data, genomics and multi-omics. Care pathways: plans, referrals and the tasks that make them real. Life outside the hospital: device readings and what the patient reports in their own words. Medical knowledge: guidelines and annotations that give the raw data clinical meaning. And underneath all of it, two things that are easy to forget until an auditor asks: identity and provenance, so we know who the data is about and where it came from, and access policy, so we know who may see it and for what purpose.

Clinical information becomes dramatically more valuable when it can be brought together and accessed together, in the same place, with its context intact. That is the model we want to support. So why is it still so hard to build?

Why the coast is hard to reach

Four obstacles stand between most healthcare organisations and that context, and I have yet to meet one that is exempt from all of them.

The first is fragmentation. Each specialised system holds a part of the patient’s story and none of them holds the whole. The laboratory system knows the results, the pharmacy system knows the dispensations, the imaging system knows the scans, and the person who knows how they fit together is usually a clinician with a very good memory.

The second is complex data. Clinical narratives, coded entries, images, device event streams and genomic files have different structures and change at different speeds. A schema that was right for encounters is wrong for wearables, and a model built for either will struggle with a genome.

The third is the semantic gap. Moving a record between two systems does not guarantee that both read it the same way. Local codes, terminologies and free text need mapping and validation, and the mapping is never finished. The honest approach keeps the original evidence available alongside the validated interpretation, so that nobody has to trust a translation they cannot check.

The fourth is rules and trust. Who may access what depends on identity, on the purpose of the access, on the patient’s rights, and on the law of the country the data sits in. In Europe that last clause is not a footnote. The same query can be legitimate in one member state and unlawful in the next.

What the technology must provide

If those are the obstacles, the requirements for a trusted data foundation almost write themselves. I framed them in four capabilities, deliberately stated as outcomes rather than as products.

Connected sources. Bring heterogeneous information about the patient together across the systems that already exist, without waiting for them to be replaced. Replacement is a decade-long programme; connection is a quarter-long one.

Clinical meaning. Keep the source evidence and its provenance, and support validated mappings on top of it. The original and the interpretation have to coexist.

Current context. Make new events available to each workflow at the speed that workflow needs. Some need milliseconds, some need minutes; none of them need yesterday.

Governed reuse. Control and trace access for care applications, for analytics and for AI, according to purpose. The same patient context will be read by a clinician, a population health dashboard and an agent, and each of them is entitled to a different slice of it.

These are the capabilities that make the longitudinal approach work in practice, rather than on a roadmap slide. In the second part of the webinar Francesc showed how a modern data platform delivers them, and the demo made them concrete: searching across a patient’s history, reacting to a new event in clinical context, and serving an authorised answer to an application that was never allowed to see the whole record.

Healthcare needs a memory

A few weeks ago I argued that AI agents do not have an intelligence problem but a memory problem: what they know, when they know it, what they are allowed to know, and what they should have forgotten. Healthcare is the sector where that thesis stops being an analogy.

A health system that cannot remember a patient across encounters, providers and years is not short of intelligence. It has excellent clinicians, increasingly good models, and more data than it can read. What it lacks is a memory: connected, current, meaningful, and governed. The five winds are already blowing. The coast is in sight. The work in front of us is the hull.