A drug performs beautifully in its clinical trials. The endpoints are met, the safety profile is clean, and the commercial team builds a forecast around the trial population. Then the drug reaches the market and behaves differently. Patients are older, sicker, and more varied than the trial cohort. They miss doses. They have conditions the trial excluded. The forecast, built on a controlled population, starts to drift from reality within the first two quarters.
Stelios Tzellos of the UK has seen this gap between trial data and real-world use catch forecasting teams off guard across the industry. As a professional in business insights, analytics, and oncology marketing at AstraZeneca, with prior roles at GlobalData and IQVIA, he treats real-world evidence not as an afterthought but as a correction to the assumptions a trial inevitably builds in.
The Trial Population Is Not the Market
Clinical trials are designed to isolate a drug’s effect. That means tight inclusion criteria, careful patient selection, and monitoring that no community oncology practice can match. The result is clean data about a narrow group. It tells you what the drug can do under ideal conditions. It tells you much less about what the drug will do once it meets the full range of patients a real clinic treats.
Forecasting teams know this in theory. In practice, the trial data is what they have when the model gets built, so the trial population quietly becomes the basis for projection. The drift starts there. A drug studied in fit patients with good organ function gets prescribed to frail patients with several other conditions, and the response rates, discontinuation rates, and treatment duration all shift.
Why Oncology Makes the Gap Worse
Oncology amplifies the problem because the stakes and the variability are both high. Treatment sequencing in the real world rarely matches the clean lines of a trial protocol. An oncologist might use a drug earlier or later than the label suggests, combine it differently, or switch therapies based on tolerance rather than progression. Each of these choices moves patient numbers in ways a trial-based model did not anticipate.
Tzellos worked on oncology disease insights at IQVIA, where the difference between trial assumptions and observed behavior showed up repeatedly. A forecast that assumed strict adherence to a treatment algorithm would overstate uptake in one line and understate it in another. The biology explained part of it. Physician habit explained the rest.
What Real-World Evidence Actually Shows
Real-world evidence comes from sources the trial never touches: insurance claims, electronic health records, patient registries, and pharmacy data. These sources are messy. They contain gaps, coding errors, and inconsistencies that frustrate anyone hoping for clean inputs. But they describe what actually happened to real patients, which is the thing a forecast is trying to predict.
The skill is in reading that messy data without drawing false conclusions from it. Tzellos approaches this the way his scientific training taught him to approach any dataset, by asking what the data can and cannot support before building an argument on top of it. His PhD work in molecular biology at Imperial College London centered on exactly that discipline: distinguishing signal from artifact.
Building Forecasts That Expect the Gap
The teams that forecast well do not treat real-world divergence as a surprise. They build models that expect it. That means presenting ranges instead of single numbers, testing how sensitive the forecast is to assumptions about adherence and patient mix, and updating projections as post-launch data arrives.
At GlobalData, Tzellos built epidemiology models for oncology and haematology indications including Hodgkin’s lymphoma, where the difference between diagnosed populations and treated populations matters enormously. At AstraZeneca, he leads cross-functional projects that bring real-world signals into product strategy rather than waiting for them to disrupt it.
The trial tells you whether a drug works. Real-world evidence tells you how it will be used, by whom, and for how long. A forecast that ignores the second question will keep getting surprised by the answer.