Check our work.

The question

A synthetic release produced under ε-differential privacy comes with a mathematical ceiling on what any attacker can learn about a real patient. How close do real, well-resourced attacks get to that ceiling, and where does the gap come from?

Why it matters

Formal guarantees are only persuasive to a hospital or an ethics board when someone independent has tried hard to break them. Published attacks on synthetic data show that similarity scores can hide real leakage. An open, reproducible attack harness is the evidence the field is missing.

What you would do

• Survey and implement the main attack families from the literature: shadow-model membership inference, distance-based attacks, attribute inference and record linkage.
• Run them against public-seed releases (CMS DE-SynPUF) at several privacy budgets, with a clear threat model for each attack.
• Compare measured attack advantage against the theoretical bound printed on the release certificate.
• Where a partner’s secure environment and ethics approval allow it, repeat the strongest attacks on releases made from real records, without data leaving the environment.

What you would produce

• An open-source attack harness with documented threat models
• A results report: where attacks come close to the bound, and where they do not
• A publishable paper; negative results count

Good fit if you have

Python, ML fundamentals, some probability. Interest in security or cryptography. Prior work with privacy attacks is a plus, not a requirement.

Suits

KTH (Cybersecurity, ML) · SU (DSV)

Download the brief (PDF) ↓ Apply for M1 →

The question

When a researcher runs a standard cardiometabolic analysis, such as diabetes outcomes or heart-failure survival, on a synthetic cohort instead of the real one, do they reach the same conclusions: the same direction of effect, overlapping confidence intervals, the same curves?

Why it matters

The point of synthetic data in research is that inferences carry over. Most evaluations report distance metrics instead of asking whether the scientific answer changes. A clear protocol for inference agreement is what a research group would accept before trusting a synthetic cohort. We start in diabetes and cardiovascular care because the conditions are common and the analyses are well established.

What you would do

• Select three to five published diabetes and cardiovascular analyses that can be reproduced on public data: risk-factor regression, Kaplan–Meier and Cox survival analysis, comorbidity prevalence.
• Re-run each analysis on the real seed data and on synthetic releases at several privacy budgets.
• Measure agreement: effect-size difference, confidence-interval overlap, sign and significance agreement, curve distance.
• Propose a compact inference-agreement report that could sit on a release certificate.

What you would produce

• A reproducibility protocol other groups can reuse
• Agreement results across analyses and budgets
• A short reporting template for inference agreement

Good fit if you have

R or Python, regression and survival analysis, careful reporting. Epidemiology or biostatistics background.

Suits

KI (Epidemiology, Biostatistics & Data Science) · SU (Statistics)

Download the brief (PDF) ↓ Apply for M2 →

The question

Synthetic data is scored with dozens of fidelity metrics. Which of them predict how well a downstream task performs, and does the answer change between software testing, cohort sizing and model training?

Why it matters

A single utility score invites the wrong reading. Buyers and access bodies need to know which evidence counts for which purpose. Mapping metrics to tasks turns fitness for purpose from a slogan into something measurable.

What you would do

• Catalogue fidelity metrics in common use: marginal divergences, correlation structure, propensity scores, train-synthetic-test-real performance.
• Define a set of downstream tasks of different difficulty on public seed data.
• Measure metric scores and task performance across many synthetic releases and look for which metrics are predictive, and where.
• Recommend a minimal, purpose-specific evidence set.

What you would produce

• A metric-to-task predictiveness study
• A recommended evidence set per purpose
• Open evaluation code

Good fit if you have

Python, statistics, experimental design. Comfortable running many experiments and reasoning about them.

Suits

KTH (ML, Data Science) · SU (Statistics, CS)

Download the brief (PDF) ↓ Apply for M3 →

The question

A synthetic cohort is only usable if it loads into the systems researchers already use. How often do synthetic records violate FHIR profiles, OMOP CDM conventions and basic clinical plausibility, and how should that be tested and reported?

Why it matters

Hospitals and research platforms run on standards. Structural errors in synthetic data waste analysts’ time and erode trust faster than statistical gaps. A standard conformance suite is a practical, reusable contribution.

What you would do

• Assemble a conformance suite: FHIR profile validation, OMOP data-quality checks (e.g. OHDSI tooling), referential integrity and clinical plausibility rules.
• Apply it to synthetic releases and to their public seed data as a baseline.
• Classify the violations found and how they affect downstream use.
• Propose a conformance section for a release certificate.

What you would produce

• An open conformance test suite
• A violation taxonomy with examples
• A reporting format for interoperability

Good fit if you have

Health data standards (FHIR, OMOP) or willingness to learn them quickly; SQL and Python.

Suits

KI / SU (Health Informatics) · KTH (Medical Engineering)

Download the brief (PDF) ↓ Apply for M4 →

The question

Under a fixed privacy budget, does synthetic data serve some cardiometabolic patients worse than others, such as younger people with type 2 diabetes, women with heart disease, or older patients with several conditions? How should that be measured and shown to the people relying on the data?

Why it matters

Published work shows that privacy noise can land hardest on small subgroups, and most medical studies do not report subgroup results at all. Regulators and the EU AI Act increasingly ask for subgroup evidence. Measuring it honestly is a design requirement for release documentation.

What you would do

• Define subgroups among diabetes and cardiovascular patients in public seed data (age, sex, region, comorbidity) and a set of subgroup-level fidelity and inference measures.
• Measure subgroup gaps across synthetic releases at several privacy budgets.
• Relate the gaps to subgroup size and to the downstream tasks affected.
• Propose a subgroup section for a release certificate.

What you would produce

• A subgroup-utility benchmark on public data
• Analysis of which groups lose most, and when
• A reporting template for subgroup results

Good fit if you have

Statistics, R or Python, interest in equity and fairness in health research.

Suits

KI (Public Health, Epidemiology) · KTH (ML)

Download the brief (PDF) ↓ Apply for M5 →

The question

When is a synthetic dataset anonymous enough, and documented well enough, to be released under the European Health Data Space? What should an access body be able to check?

Why it matters

EHDS access bodies will have to approve or refuse synthetic outputs, yet there is no agreed standard for what adequate means. GDPR anonymity, the AI Act’s data-governance rules and EHDS secondary-use rules each pull in a different direction. A clear set of criteria is useful to regulators and to every producer.

What you would do

• Map the relevant provisions: GDPR anonymisation, EHDS secondary use, AI Act data governance, relevant EDPB guidance.
• Analyse how formal privacy guarantees and empirical risk assessments fit those provisions.
• Compare approaches in two or three member states or authorities.
• Propose checkable adequacy criteria, including what a release should state it must not be used for.

What you would produce

• A doctrinal and comparative analysis
• Proposed adequacy criteria for access bodies
• A policy brief suitable for publication

Good fit if you have

EU law, data protection law. Comfort reading technical summaries; no programming needed.

Suits

SU (Law & IT, European law) · KI (Health policy)

Download the brief (PDF) ↓ Apply for M6 →

The question

What does it take for a diabetes or cardiology clinic to host a synthetic-data validation, so that raw patient records stay inside its environment and only synthetic data and evaluation results come out?

Why it matters

The strongest evidence for synthetic data comes from real clinical records. We start in cardiometabolic care, at a diabetes or cardiology clinic, with oncology as a later stage. Hospitals will only take part if governance is clear: ethics approval, DPIA, roles, access controls and exit criteria. A tested protocol lowers the barrier for the first sites.

What you would do

• Review existing secure-processing models (e.g. trusted research environments, MONA-style remote access) and Swedish requirements.
• Interview data stewards, ethics and IT security staff about their conditions for participating.
• Draft a step-by-step validation protocol: approvals, roles, what runs inside, what may leave, stop criteria.
• Pilot the protocol end-to-end on public cardiometabolic data in a simulated environment.

What you would produce

• A governance and validation protocol hospitals can review
• Interview findings on adoption barriers
• Checklists and templates for a first site

Good fit if you have

Health informatics or health systems background; interviewing and qualitative analysis; basic technical literacy.

Suits

KI (LIME, Health Informatics) · KTH (Health systems)

Download the brief (PDF) ↓ Apply for M7 →

The question

Where in the drug-development pipeline for cardiometabolic disease can certified synthetic data be used today, and what evidence would a pharma team, an ethics board or a regulator need at each stage?

Why it matters

Synthetic data is discussed for discovery, trial feasibility and synthetic control arms, but expectations differ sharply by stage. EMA and FDA statements give partial guidance. A clear stage-by-stage evidence map helps companies and research institutes decide where to start.

What you would do

• Pick one cardiometabolic therapy area, such as type 2 diabetes or heart failure, and map its stages: target discovery, cohort and feasibility analysis, trial design, external comparators, post-market.
• Review EMA, FDA and Nordic guidance and published use cases at each stage.
• Interview people in Nordic pharma, CROs and research institutes about what they would accept.
• Build an evidence ladder: which uses are ready, which need more validation, which are not appropriate.

What you would produce

• A stage-by-stage evidence map
• Interview insights from industry
• Recommendations for early adoption pathways

Good fit if you have

Life-science background; interest in regulation and business; interviewing and synthesis.

Suits

KI (Bioentrepreneurship, Regulatory science) · SU

Download the brief (PDF) ↓ Apply for M8 →

The question

Can we build audit methods that let an independent party verify, with statistical confidence, that a synthetic clinical release behaves as its stated privacy guarantee says, without seeing the producer’s internals?

Why it matters

Privacy auditing is an active research area, but clinical synthetic data has its own structure, such as linked tables and coded clinical events. Third-party verifiability is what turns a vendor claim into a trust standard.

What you would do

• Develop and evaluate black-box audit methods for synthetic tabular and relational clinical data.
• Study how tight empirical lower bounds on privacy loss can be made for realistic data.
• Validate on public data and, where approved, inside partner secure environments.
• Publish open audit tooling.

What you would produce

• A research programme of three to four papers
• Open audit tooling
• Collaboration across academia, hospitals and industry

Good fit if you have

Strong background in ML, statistics or security; master’s degree in a relevant field.

Suits

KTH · SU · co-supervised

Download the brief (PDF) ↓ Apply for P1 →

The question

What body of evidence should be required before synthetic clinical data is used for a given research purpose, and how can that evidence be standardised so registries, access bodies and regulators accept it?

Why it matters

There is no reporting guideline for synthetic health data comparable to those for trials or observational studies. Sweden’s registers and secure environments make it an unusually good place to build and test one.

What you would do

• Systematically review evaluation practice and existing reporting guidelines.
• Run reproducibility studies on public and, where approved, register-based data.
• Develop and test a reporting guideline through consensus methods with clinicians, statisticians and regulators.
• Study adoption in real research projects.

What you would produce

• A proposed reporting guideline
• A series of reproducibility studies
• Evidence on adoption and acceptance

Good fit if you have

Epidemiology or biostatistics master’s; interest in research methods and regulation.

Suits

KI · co-supervised

Download the brief (PDF) ↓ Apply for P2 →