The health-data questions I answer for

Health data is a special category under the GDPR, and pseudonymised data remains personal data, so every research use of our records needs a lawful basis, a purpose patients can understand and safeguards proportionate to the risk. In France that also means working within the CNIL's reference methodologies and its framework for health-data warehouses, informing patients and honouring their right to object. At Aurevalle I designed the process that turns those obligations into a working routine: one data-access committee of clinicians, an ethicist and a patient representative, with the data protection officer as a non-voting adviser; standard agreements for academic and industry partners; a public register of every project; and an objection register checked automatically at every query. The committee approved 64 projects in 2025. Speed is a welcome side effect, since the median request now takes eleven weeks rather than seven and a half months, but the purpose is different: I can explain any use of a patient's record to that patient.

Our warehouse draws on the electronic health record, the laboratory and pathology systems, imaging metadata, pharmacy dispensing and the coded discharge data that French hospitals produce for activity-based payment. It covers 3.1 million patients from 2012 onwards. The engineering was the easier part. The harder part was documenting how each source variable came into being: who records it, at what moment, for what purpose, and when its meaning changed, as several meanings did when two of our hospitals moved to a new record system in 2019. Every table now carries that provenance, written by data stewards who spend part of each week with clinical teams. I fund stewardship before analytics, and 14 of our 86 staff are stewards. When a researcher asks why a variable behaves strangely, the answer is usually already in a steward's note rather than waiting in a bug report.

A CE mark under the EU Medical Device Regulation tells me that an algorithm conforms to the regulation for its intended use. It does not tell me how it performs on our scanners, our patients and our radiographers' habits. Since 2024 every imaging algorithm we consider runs silently on our images first, its output hidden from clinicians and compared with radiologists' reports, with a panel adjudicating disagreements. Of five tools assessed so far we deployed three, restricted one to the settings where it performed well and declined one. After go-live we monitor monthly: flag rates by site and scanner, radiologist overrides and an audit of a sample of films the tool passed as normal. That monitoring caught a fall in specificity after a scanner replacement at Arras in November 2025. I am careful about what this work shows: silent running measures the algorithm, not how clinicians behave once they can see its output.

At Tornvig Pharma I led the studies that turned routine data into evidence for medicines regulators and health technology assessment bodies: two post-authorisation safety studies required by the European Medicines Agency, comparative analyses for reimbursement dossiers, and natural-history work for a rare-disease medicine. I learnt that the decisive question is rarely statistical sophistication. It is whether the data were recorded in a way that can answer the specific question, with known gaps, consistent definitions and measurements taken often enough. I introduced a written feasibility assessment before every protocol, which in 2019 led me to withdraw an external control arm from a dossier. Now that I sit on the hospital side, I apply the same test to the requests we receive and write a short fitness note for each one, so that a sponsor knows what our records can bear before it builds a study on them.

Ten years at Orsinval Clinical Research taught me how trial data is protected and how it fails. I was lead statistician on eleven trials and then directed a European data-management and biostatistics group of 170 people responsible for 40 to 50 concurrent studies. The international good clinical practice guideline encourages risk-based monitoring, including central review of accumulating data, and in 2014 that approach exposed fabricated visit data at one site in a phase III trial before the database was locked. From 2015 we ran statistical checks on every phase II and III study from the first patient: terminal-digit preference, implausibly low variability, patterns repeated across patients and anomalies in visit dates. I now bring that discipline to our hospitals as investigator sites. Our trial units receive the same anomaly reports on their own data that sponsors see, because a site should never be the last to learn that its records look wrong.