Stelios Tzellos | Epidemiology Models Aren't Predictions. They're Arguments.

Stelios Tzellos

Every pharmaceutical company has an epidemiology model for its key products. Most of those models produce a single number: the estimated patient population in a given market for a given indication. And most of those numbers are wrong. Not because the math is bad, but because the assumptions behind the math are treated as facts rather than what they actually are: arguments about how a disease behaves in a population.

Stelios Tzellos of the UK understands this distinction. As a pharmaceutical analytics professional with a background in molecular biology and oncology strategy, Tzellos has built, evaluated, and challenged epidemiology models across multiple organizations. The ones that hold up over time share a common trait: their creators understood what they were arguing, not just what they were calculating.

Where the Numbers Come From

An epidemiology model in oncology starts with incidence data: how many new patients are diagnosed each year with a particular cancer. From there, it layers on prevalence, survival rates, treatment eligibility, biomarker prevalence, and line-of-therapy distribution. Each layer adds assumptions, and each assumption is a choice.

At GlobalData, Stelios Tzellos developed epidemiology models for oncology and haematology indications, including work on Hodgkin's lymphoma and other cancer types. Building those models meant deciding which data sources to trust, how to handle inconsistencies between registries, and how to project changes in diagnosis rates over time. None of those decisions are purely mathematical. They require judgement.

The difference between a useful model and a misleading one often comes down to whether the modeller documented those judgements and defended them. A model that says "the eligible patient population for this indication is 45,000" is only useful if someone can explain why it's 45,000 and not 38,000 or 52,000.

The Molecular Biology Connection

Tzellos completed his PhD in Molecular Biology at Imperial College London, studying Epstein-Barr virus gene regulation. That training taught him something that many analysts miss: biological systems are variable by nature. The same virus behaves differently in different cellular contexts. The same cancer presents differently across patient subgroups.

When you bring that mindset to epidemiology modelling, you're less likely to treat a single estimate as the answer. You're more likely to build ranges, test sensitivities, and present results as a spectrum of scenarios rather than a point prediction. That approach produces less tidy slides but more honest analysis.

How This Plays Out in Practice

At IQVIA, Stelios Tzellos worked in oncology disease insights and the Analytics Center of Excellence. He worked with pharmaceutical clients on forecasting and commercial planning, where epidemiology models feed directly into revenue projections and investment decisions.

The clients who got the most from his work were the ones who treated the model as a starting point for discussion rather than a final answer. When a client pushes back on an assumption, that's not a failure. That's the model doing its job: surfacing the questions that need to be answered before a commercial decision can be made.

At AstraZeneca, where Tzellos now works across business insights, analytics, and oncology marketing, this approach is part of how portfolio decisions get made. The models don't tell leadership what to do. They tell leadership what to debate.

Building Better Models Means Asking Better Questions

The pharmaceutical industry spends significant resources on epidemiology modelling, and a good portion of that investment is wasted on false precision. A model that produces a number with three significant figures when the underlying data supports one is worse than useless. It creates false confidence.

Stelios Tzellos has built a career around the idea that the value of a model lies in the questions it raises, not the answers it gives. His training at Imperial College London, his consulting work at GlobalData and IQVIA, and his current role at AstraZeneca all reinforce the same point: the best analysts don't just build models. They build arguments.

If your epidemiology model doesn't make someone uncomfortable, it probably isn't telling you anything you didn't already believe.