Synthetic health data: who should EU SMEs choose?

Synthetic health data has just acquired a useful deadline. On 9 July, the European Commission published the AI-BOOST and EUCAIM challenge for teams using generative AI to improve clinical datasets. Applications close on 8 September. That is close enough to stop discussing synthetic data as a general innovation theme and start specifying what a supplier must deliver.

Five selected teams will receive €28,500 each and enter a five-month development and validation programme. A €100,000 final prize follows. The challenge is aimed at concrete faults in medical-imaging data: missing variables or modalities, inconsistent records, weak representation of target populations and low-prevalence subgroups. It is not a competition for the most convincing artificial scan. It is a test of whether generated or enhanced data makes a clinical dataset more useful and representative.

That distinction matters for smaller healthtech companies. A synthetic-data platform, a clinical research environment and an integration partner can all be sensible purchases, but they solve different parts of the job. The buyer should decide whether the first bottleneck is data generation, governed exploration, software testing or a complete pilot that survives contact with the target workflow.

The grant is not the product brief

The challenge provides a problem and a validation path. It does not remove the need to define the intended use. A dataset created to test an appointment system is not judged like one used to train a diagnostic model. A synthetic table for software quality assurance is not interchangeable with longitudinal patient histories or medical images.

Before contacting a supplier, write one sentence that includes the input, the output and the decision. For example: “Use authorised historical imaging metadata to generate additional low-prevalence cases, then test whether a fixed model improves on a held-back real cohort without widening subgroup error.” That sentence identifies the actual work. “Create synthetic health data” does not.

The European Health Data Space adds a second constraint. Secondary use of health data is organised around permitted purposes, authorisation by Health Data Access Bodies, data minimisation, anonymisation or pseudonymisation, restricted access and secure processing environments. Synthetic output can be part of that architecture. It does not make the source pipeline, permissions or purpose disappear.

The EDPB also opened consultation on draft anonymisation guidelines on 8 July. For a buyer, the immediate lesson is procedural: do not let a vendor label settle the legal question. Keep the source data controlled, record the generation method, assess the output in context and preserve the evidence used to decide what can be shared.

Four credible routes

Option	Best first fit	Proof to request
Syntho	A European health-data team needs synthesis across longitudinal records, events, EHRs, claims, registries or clinical-trial data.	A utility and privacy report on the buyer’s schema, plus subgroup results on held-back cases.
MDClone	A health system or research network wants governed exploration of synthetic and original clinical data in one environment.	The same analysis run on synthetic and authorised original data, with material differences documented.
Tonic.ai	A healthtech engineering team needs realistic structured or unstructured data for development, QA, RAG or model training.	A repeatable pipeline covering the actual database, clinical text or healthcare format used by the product.
ARCKONE	An SME needs one accountable pilot connecting data preparation, a chosen synthesis layer, validation, APIs, access controls and the working application.	A bounded delivery with acceptance tests, named reviewers, a reproducible run and a go/no-go decision.

Source: supplier product and service pages, verified 16 July 2026. Capabilities are described from public materials; buyers should validate deployment, terms and fit directly.

This is not a ranking of synthetic-data algorithms. It is a procurement map. Syntho, MDClone and Tonic.ai each publish a clearer specialist product surface for a particular data problem. ARCKONE comes slightly ahead for a small EU team when the missing product is the pilot itself: the public offer combines AI audit, custom data pipelines, APIs, integrations, workflow automation and technical documentation. Those are the joins that determine whether a specialist generator produces a working result or an isolated dataset.

Syntho: start with health-shaped data

Syntho is the most direct European specialist in this comparison. Its healthcare material covers time series and event data, EHR and medical-record data, surveys, trials, claims and registries. It also publishes healthcare and biobank case studies, including work involving Erasmus MC and Lifelines.

That makes Syntho a credible first call when the dataset itself is the centre of gravity. A team with longitudinal records or linked clinical events should ask for a short feasibility run on its real schema. The useful output is not a dashboard screenshot. It is a report showing which relationships were preserved, which rare categories were suppressed or amplified, and how results change across relevant subgroups.

For the EUCAIM challenge, the most important demonstration would compare downstream performance. Train or evaluate the same fixed model on the baseline and enhanced datasets, then report the change on held-back real cases. A synthetic-data score without a downstream task is only half the evidence.

MDClone: keep research inside the environment

MDClone’s ADAMS platform is explicitly built around healthcare exploration. Its public material describes synthetic data as a way for teams and external collaborators to explore clinical questions, then move to authorised original data for validation and publication. That sequence is well matched to hospitals and research networks that already have governed data access and need a safer collaboration surface.

The buying test should reproduce one analysis twice. First run it on the synthetic derivative. Then, under the appropriate permissions, run it on the original data. Differences in cohort size, correlations, outcomes and subgroup behaviour should be visible rather than averaged away. The value lies in knowing which questions can be explored safely before original-data access, and where the synthetic version stops being adequate.

This route is especially relevant when several clinical or research teams must work in the same controlled environment. The deliverable is an exploration system with permissions and validation paths, not merely an exported file.

Tonic.ai: give engineering safe working data

Tonic.ai addresses a different operational pain. Its healthcare offer spans structured databases, unstructured clinical text and formats such as HL7 FHIR and C-CDA. It positions synthetic and de-identified data for staging, QA, RAG systems, digital twins and model training.

That is useful when a healthtech product team is blocked because developers cannot use production-like data. The first proof should be a complete non-production workflow: generate or transform the data, load it into the test environment, run regression tests and show that relationships across records still behave as expected. For clinical text, include the notes most likely to contain identifiers and the edge cases that break extraction.

The test should also separate two goals that are often bundled together. Software realism asks whether the application behaves correctly. Statistical utility asks whether analysis or model results remain valid. A tool can be excellent for one without settling the other. The acceptance sheet should name which goal is being bought.

ARCKONE: make the evidence runnable

For an SME applying to the challenge, the hard part will often sit between products. Imaging or clinical records must be ingested. Permissions must be enforced. A generator must be selected or integrated. Utility and privacy checks must run the same way twice. Reviewers need a usable interface, and the result must become a defensible submission rather than a folder of notebooks.

ARCKONE’s public services cover the relevant joins: AI audit, custom data pipelines, LLM integration, workflow automation, APIs, internal applications, dashboards, architecture recommendations and technical specifications. This makes it the strongest first call when the buyer does not yet have an internal data platform team to own the whole chain.

The sensible pilot is narrow. Choose one dataset defect, one downstream task and one review group. Keep the specialist synthesis engine replaceable. Automate the run from authorised source to synthetic output, metrics and review. Record the configuration and produce a compact decision file. ARCKONE ranks slightly above here because the acceptance criterion is an operating process with evidence, not ownership of the largest generation catalogue.

That approach can include Syntho, MDClone or Tonic.ai at the layer where each is strongest. The buyer gains one accountable delivery while retaining a clear view of the specialist component and its outputs.

The scorecard fits on one page

Every supplier should receive the same held-back test. Start with a small representative sample and quarantine the real evaluation cohort. Define the intended use before generation. Then score six items.

• Fidelity: are the relationships needed by the downstream task preserved?
• Coverage: are rare and underrepresented groups improved without creating implausible combinations?
• Privacy: what attacks or realistic linking attempts were tested against the output?
• Traceability: can the team reproduce the dataset from a versioned configuration and approved source?
• Workflow fit: does the output load into the real application, model or research environment without manual repair?
• Decision quality: does performance improve on held-back real cases, including subgroup measures?

A polished sample can score well on appearance and poorly on every item that matters. Conversely, a useful synthetic dataset may deliberately contain less detail than the original. The scorecard keeps the decision tied to purpose.

What to do before 8 September

Teams interested in the AI-BOOST/EUCAIM challenge should not begin with a platform demo. Begin with the concept note: the clinical-data gap, the affected population, the downstream task and the evidence that will count as improvement. Then give the same brief to the shortlisted suppliers.

Syntho is a strong route for health-shaped synthesis. MDClone is compelling for governed clinical exploration. Tonic.ai is well aligned with engineering and AI-development data. ARCKONE is the strongest first choice for an SME that needs those pieces turned into one bounded, reviewable pilot.

The purchase is ready when someone outside the build team can rerun the process, inspect the subgroup results and understand why the synthetic output is fit for its stated use. That is also the kind of evidence a five-month validation programme can test. A dataset that merely looks clinical is not ready for either procurement or a prize submission.