Six roles on three sides of the patient record

Mandate

Give the group one accountable data function: build a clinical data warehouse clinicians would trust, put secondary use of patient data on a lawful and transparent footing, and bring imaging AI under a validation discipline.

Outcomes

• Opened a clinical data warehouse in 2023 covering 3.1 million patients' records since 2012, built to the CNIL's framework for health-data warehouses.
• Replaced row-level extracts with an analysis environment in April 2024; 38 external projects ran in it in 2025 and partnership income reached €2.6 million.
• Validated five imaging algorithms on local images before use: three deployed, one restricted to adult upright films, one declined.
• Cut the median time from research request to approved access from seven and a half months to eleven weeks.
• Introduced a public project register and an objection register enforced in every query; 3,140 objections recorded since 2023.

Decisions in context

I joined Aurevalle in March 2021 as its first chief data officer, reporting to the directeur général and sitting on the executive committee. The group had a decade of electronic records spread across several systems, a research department that had been sending extracts to partners for years, and imaging and laboratory services under staffing pressure. My mandate was to give the group one accountable data function. I built the warehouse first, opening it to approved research in 2023, then the analysis environment in 2024, and in parallel a validation discipline for imaging algorithms. I lead 86 people with six direct reports and manage a €9.8 million operating budget and a capital programme of €14 million that ran from 2022 to 2025. The data protection officer is independent of me. The work I value most is the least visible: the stewards' notes that explain what each variable means and when its meaning changed.

Reason for the move

Current role. Not seeking an executive move; considering one non-executive seat with the consent of the directeur général and the board.

Mandate

Build a real-world evidence function that regulators and health technology assessment bodies would accept, starting with the post-authorisation safety commitments for the company's respiratory biologic.

Outcomes

• Delivered two post-authorisation safety studies required by the European Medicines Agency on their agreed schedules.
• Withdrew an external control arm from a rare-disease reimbursement dossier and funded standardised registry data collection at 11 centres.
• Made a written feasibility assessment compulsory before every real-world study protocol from 2020.
• Reduced data licences from 14 to 6, saving €1.9 million a year.

Decisions in context

I moved to Copenhagen in September 2016 to build Tornvig's real-world evidence and epidemiology function, which until then had been a small group buying data licences study by study. Tornvig is listed in Copenhagen and controlled by a foundation, and its leadership took a long view of evidence that I came to value. I reported to the senior vice president for global medical affairs and grew the team to 34 people in Copenhagen and Paris. We delivered two post-authorisation safety studies required by the European Medicines Agency for the company's respiratory biologic, supported reimbursement dossiers in more than a dozen countries and cut our data licences from fourteen to six. The external control decision in 2019 was the defining moment of those years, and the feasibility standard that followed it was my most durable contribution. I left in February 2021 to work on the hospital side of the data.

Reason for the move

Wanted to work where patient data is created and governed rather than where it is bought, and to return to France.

Mandate

Bring two acquired data-management teams onto one process and one electronic data capture standard, and make database locks reliable.

Outcomes

• Cut the median time from last patient visit to database lock from 9.5 weeks to 5.5.
• Merged the Paris and Warsaw teams onto a single data capture standard and one set of procedures.
• Identified fabricated visit data at one investigator site in a phase III trial before lock in 2014.
• Introduced central statistical monitoring on every phase II and III study from 2015.

Decisions in context

In January 2011 I became director of clinical data management and biostatistics for Europe at Orsinval, which had grown by acquiring two smaller data-management firms, one of them in Warsaw. The teams used different data capture standards and different definitions of a clean database. I merged them onto one standard and one set of procedures and reduced the median time from last patient visit to database lock from 9.5 weeks to 5.5. I led 170 data managers, programmers and biostatisticians, with seven direct reports, responsible for 40 to 50 concurrent trials. The fabrication case in 2014 changed my view of what data management is for: not only clean data but credible data. From 2015 we applied central statistical monitoring to every phase II and III study. In 2016 a sponsor-side role in real-world evidence offered the questions I most wanted to work on next.

Reason for the move

Recruited to build a sponsor-side real-world evidence function, the work I wanted to do next.

Mandate

Write statistical analysis plans and lead analysis for cardiovascular and renal trials, and represent the statistics function in sponsor meetings.

Outcomes

• Lead statistician on eleven phase II and III trials, five of them pivotal.
• Wrote Orsinval's first standard for blinded data review before unblinding.
• Trained fourteen statisticians, several of whom now lead teams.

Decisions in context

I joined Orsinval in April 2006 as a senior biostatistician and became biostatistics lead for cardiovascular and renal trials in 2008, with a team of eight. I wrote statistical analysis plans, led analyses and represented the company in meetings with sponsors and, occasionally, with the regulators they had invited. I was lead statistician on eleven phase II and III trials, five of them pivotal. I also wrote Orsinval's first standard for blinded data review, which defined what a team could examine before unblinding and recorded every decision taken along the way. After the registry work in Brest, trials taught me the value of defining every variable before the first patient is enrolled, and the cost in effort and money of doing so properly. I trained fourteen statisticians in those years, and I still hear from several of them when a protocol puzzles them.

Reason for the move

Promoted to lead clinical data management and biostatistics for Europe.

Mandate

Link the centre's tumour registry with its treatment and follow-up records so that outcomes could be analysed by treatment.

Outcomes

• Linked registry and treatment records for about 18,000 patients.
• Published survival analyses for breast and head-and-neck cancers.
• Wrote the centre's first data dictionary recording who created each variable and when.
• Led an audit of pathology coding that corrected about 6 per cent of tumour classifications.

Decisions in context

At the Kerverzic cancer centre in Brest I was the epidemiologist in a small unit linking the centre's tumour registry with its treatment and follow-up records. We linked about 18,000 patients and published survival analyses for breast and head-and-neck cancers. The survival discrepancy described in my biography came from this work: three legitimate definitions of the date of diagnosis, used by three groups of people for three purposes. I wrote the centre's first data dictionary that recorded who created each variable and when, and I led an audit of pathology coding that corrected about six per cent of tumour classifications. Those three years gave me the habit I rely on most: asking how a number was produced before asking what it says. I left in 2006 because I wanted to learn how prospective trials are designed and run.

Reason for the move

Wanted to learn how prospective trials are designed and run, which a contract research organisation could teach quickly.

Mandate

Analyse reimbursement claims to support pricing and the prevention programmes the mutual was beginning to offer its members.

Outcomes

• Built the mutual's first member-level claims dataset for people with diabetes.
• Showed that fewer than half of members with diabetes had an annual eye examination reimbursed, which shaped the first prevention programme.
• Produced quarterly claims indicators used in pricing reviews.

Decisions in context

My first post after my master's degree was as a statistician at Mutuelle Hardivel, a complementary health insurer in Paris with about 900,000 members. I analysed reimbursement claims for pricing and for the prevention programmes the mutual was beginning to offer. I built its first member-level claims dataset for people with diabetes, which showed that fewer than half of them had an annual eye examination reimbursed, and the finding shaped the mutual's first prevention programme. Claims taught me two lasting lessons: a payment system records a great deal about care, and it is silent about the reasons for it. I also learnt to explain uncertainty to actuaries and managers who wanted a single number and deserved an honest range. I left at the end of 2002 because I wanted to work with clinical records rather than with what they cost.

Reason for the move

Wanted to work with clinical records rather than with what care cost.