Data migration is not emptying one bucket of data into another

When migrations go wrong

In Sweden, a new regional electronic health record system in Västra Götaland was contracted in 2018 and finally went live in late 2024, after years of preparation, integration and data migration, only to be stopped three days later due to data quality problems and patient safety concerns.

In Finland, a large-scale programme to replace tens of legacy health and social care systems with a single, unified record across the Helsinki region resulted in formal complaints from clinicians, who warned that the system “threatens patient safety” because critical information is hard to find or appears to disappear in practice. The way legacy data and workflows were mapped turned the project itself into a source of new safety risks.

In the United Kingdom, Manchester University NHS Foundation Trust undertook what became one of Europe’s biggest electronic record go-lives, starting with more than 1,000 separate clinical information systems. The readiness programme alone took two years, while legacy platforms are still being decommissioned today, underlining how much effort it takes simply to understand, migrate and safely retire decades of fragmented clinical information when moving to a single modern EHR.

Germany’s costly legacy data challenge

In Germany, digital health initiatives have to work around a deep-rooted legacy infrastructures and routines. General practitioners still rely heavily on telephone, fax and postal mail to exchange clinical information, even as new systems are introduced (The Role of the Installed Base in Information Exchange Among General Practitioners in Germany: Mixed Methods Study. Journal of Medical Internet Research).

National analyses from Fraunhofer ISI. E Health in Germany, find that decades of fragmented technology infrastructure, diverse practice systems and paper-based records have left Germany with one of Europe’s more expensive but comparatively under-digitised health systems. The main challenge is not the lack of new digital tools, but the difficulty of migrating and integrating data and workflows that are deeply embedded in legacy systems.

The scale of Germany’s major data-integration programme illustrates this challenge. The Medical Informatics Initiative (Medizininformatik-Initiative, MII), coordinated by the Federal Ministry of Research, Technology and Space (BMFTR; formerly BMBF), has provided more than €400 million to its four consortia, their Data Integration Centers (Datenintegrationszentren), and cross-consortium use cases across two funding phases from 2018 to 2026. Including the “Digital Progress Hubs for Health” programme, which tests the transfer of these solutions into routine regional care, total funding reaches around €500 million. A significant part of the funding has gone towards establishing and operating the Data Integration Centers that are responsible for harmonising, extracting, and consolidating routine clinical and research data across university hospital sites.

These are not isolated examples; they are part of a pattern documented for decades. A 2023 peer-reviewed study in Government Information Quarterly, conducted under the European Commission’s ISA2 programme, found that legacy systems embed institutional knowledge, business processes and regulatory compliance in ways that make them quite difficult to replace. The core challenge is not building new systems, but safely consolidating and migrating the data, workflows, and compliance obligations embedded in the existing systems.

The complexity beneath the surface

Legacy data is the hidden blocker behind almost every major healthcare IT initiative: analytics programmes that cannot access complete patient histories; AI applications that require structured, standardised data they cannot reliably obtain; regional interoperability projects that stall because each organisation’s data looks different from the next. The real issue is that it is not just a technical inconvenience, but a problem that can’t be resolved using a standard approach.

Large hospitals often operate hundreds of clinical and administrative applications. Together, these systems contain vast numbers of technical data fields representing thousands of clinically meaningful data elements. Before a migration can begin, organisations must understand what those fields represent, how they relate to one another, and how they should map to modern standards such as openEHR or HL7 FHIR. Typically, there is no shared data governance that guarantees alignment between departments and systems.

So, the hardest part of migration is not moving data technically from one place to another. It is preserving what that data actually meant at the point of care. Field names are often opaque, business logic is undocumented, and clinical meaning is embedded in local usage, free text and workflows that have evolved over decades. As soon as organisations begin to migrate, they expose years of accumulated inconsistency and data-quality issues.

The traditional response has been to address this with specialist human effort: experienced integration teams who manually analyse legacy systems, map data to modern standards like openEHR or HL7 FHIR, and build the transformation pipelines that make data usable. This works. But it is slow, expensive, difficult to replicate, and too dependent on individuals. If those specialists leave, their knowledge goes with them.

Where AI can – and cannot – help

Artificial intelligence is increasingly presented as the answer to healthcare’s data challenges. In some respects, this optimism is warranted. In others, it reflects a misunderstanding of where the real problem lies.

AI is genuinely powerful at the analytical tasks that make legacy integration so time-consuming: reading and interpreting database schemas, identifying relationships between data fields, building structured models of complex legacy systems, and generating initial mapping proposals. Given the sheer number of 1000s of tables, clinical concepts, custom forms and valuesets in a typical hospital information system, the use of AI is very helpful to cartograph the data jungle. Work that previously required weeks of specialist analysis can, with well-designed AI tooling, be completed in days.

But AI should not be making decisions about clinical data flows. There is also always the chance that the full clinical context lives in the collective mind of the organisation in the form of tacit knowledge that just cannot be inferred from documentation, databases, and form configurations. Hence, the responsibility must remain with human specialists who understand the clinical meaning of the data they are handling. Clinical data that flows into analytics platforms, AI applications or patient care systems must be accurate, reproducible, and fully auditable.

Letting an AI system generate mappings and execute it without human oversight would introduce a level of risk healthcare organisations should not take on. It is also precisely the type of risk the EU AI Act aims to prevent, by mandating human oversight, traceability and robust risk management.

The right model is a collaboration: AI provides the analytical heavy lifting and thereby frees the time of specialists to provide the clinical and technical judgement, and execution happens through a deterministic, transparent engine in which AI plays no part. This is the only approach that can be both fast enough to be useful and safe enough to be trusted in a clinical environment.

Sources
_{Irani Z, et al. The impact of legacy systems on digital transformation in public administration. Government Information Quarterly. 2023. link}

_{Manchester University NHS FT. 1000+ legacy systems, 10 hospitals, one electronic patient record. PubMed Central. link}

_{Yle News. Doctors file complaint with health watchdog over Apotti data system. link}

_{The Register and Heise Online. Coverage of Millennium rollout and halt in Västra Götaland. link}

_{Fraunhofer ISI. E‑Health in Germany. link}

_{JMIR. The Role of the Installed Base in Information Exchange Among General Practitioners in Germany: Mixed Methods Study. link}

_{Massachusetts Institute of Technology, Secondary Analysis of Electronic Health Records. link}