Learning whose habit each number records

In 2004, at a cancer centre in Brest, I spent three months trying to understand why our five-year survival estimate for breast cancer moved by almost four percentage points depending on who ran the analysis. Nobody had made an error. The registry recorded the date of diagnosis as the date of the biopsy; the treatment records used the date the pathology report was signed; the oncologists' letters used the date of the first consultation. Each was right for its author's purpose. That was the first time I understood that a variable is a habit, and that before I count anything I need to know whose habit I am counting.

I grew up in Amiens, the daughter of a pharmacist and a mathematics teacher, and I chose statistics because it seemed to sit between my parents' two worlds. I qualified as a statistics and economics engineer at the Institut Varennier de statistique appliquée in 1999 and completed a master's degree in epidemiology and biostatistics at the Institut Marsaguet de santé publique the following year. My first post, at Mutuelle Hardivel in Paris, was analysing reimbursement claims for a complementary health insurer. Claims taught me how much a payment system can reveal about care, and how much it hides. A claim records that something was paid for. It rarely records why.

I moved to the Kerverzic cancer centre in Brest in 2003 to work with clinical records, and the survival puzzle followed. In 2006 I joined Orsinval Clinical Research, a contract research organisation in Paris, as a biostatistician on cardiovascular and renal trials. Trials were a revelation after registries: every variable defined in advance, every visit scheduled, every value checked against its source. I also learnt how much that discipline costs, and that a trial can be perfectly documented and still wrong. In 2014, as director of data management and biostatistics for Europe, I recommended delaying a database lock because one high-recruiting site's blood pressure readings looked too regular to be real. The audit that followed found fabricated visit data for 41 patients. We lost a milestone payment and, for a season, some of the client's warmth, and we gained a method of central statistical monitoring that Orsinval applied to every later trial.

In 2016 I moved to Copenhagen to build a real-world evidence function at Tornvig Pharma, a research-based company with respiratory and rare-disease medicines. I had spent ten years making data fit a protocol; now I had to judge data collected for other purposes and decide whether regulators and health technology assessment bodies would trust it. The hardest decision of those years was withdrawing an external control arm from a rare-disease dossier, because the registry behind it measured lung function too rarely to support the comparison. Our head of market access thought I had weakened the company's position for the sake of method. In two countries he was right about the consequence.

Copenhagen also showed me what a hospital can and cannot offer a sponsor. I signed data licences with hospitals across Europe and learnt to read their data dictionaries for what they left out. In 2021 Aurevalle's directeur général asked me to come back to France and sit on the other side of those agreements. The group had a decade of electronic records, a research department that wanted a warehouse, an imaging service under pressure and nobody accountable for how patient data was reused. I accepted because I wanted to work where the data is created rather than where it is bought.

At Aurevalle I lead 86 people in data engineering, warehouse operations, biostatistics, imaging informatics and data stewardship. The group's data protection officer reports directly to the directeur général and is independent of my team, which I regard as a protection for me as much as for patients. I sit on the executive committee, present to the foundation's board whenever data is on its agenda and report twice a year to its audit and risk committee. Clinical decisions belong to clinicians and medical judgements to the medical committee. My task is to make sure that the records they rely on, and the research they permit, can be explained to the patients those records describe.

I ask three things of my team. Write down how each variable was born before anyone analyses it. Tell a requester what our data cannot show as plainly as what it can. And bring me a near miss on the day it happens, not on the day it becomes an incident. We have not always met those standards. In 2022 an extract reached the wrong person at a university partner, and we notified the CNIL within the legal deadline. That episode did more than any policy paper to change how we share data.

Among French hospital data teams I am probably known for three positions that some colleagues find inconvenient: that a hospital should measure an imaging algorithm on its own films before trusting it, that pseudonymised extracts should become the exception rather than the routine, and that no dataset is research-grade in general. I hold them because I have seen the alternatives from the sponsor's chair. I am now preparing to take the same questions into a boardroom as a non-executive director, where the data decisions belong to others and my task would be to ask about them well.

Open to Board appointments · Speaking invitations

Languages / French · English · Danish