Introduction

Predictive modelling and machine learning have become central to biomedical research, and they hold real promise for delivering biomarkers that change clinical practice and public health [1, 2, 3, 4].

For many researchers, though, “machine learning” remains a slightly mysterious term — more familiar from science fiction than from their own work. This is odd, because the same researchers routinely use mixed-effects models, navigate complicated analysis pipelines, and think carefully about confounding. The obstacle is not mathematical sophistication. It is that predictive modelling asks a different question from the one their training prepared them for.

Who this book is for

You, if you are comfortable fitting a linear model and interpreting a p-value, and would like to know what changes when the goal becomes prediction rather than inference.

The aim of this book is to demystify predictive modelling by showing that its simplest — and already very powerful — forms are as accessible as the tools you use every day. Not by simplifying the ideas, but by being concrete about them.

How it works

Everything is illustrated on one dataset, with python code, in a single running example: predicting age from brain structure. The data is structural MRI, but no neuroimaging knowledge is assumed. If you would like some background anyway, there are many good resources available free online.

Practice pages are interactive notebooks. They run on your own machine or, with one click, in the cloud — look for the rocket icon at the top of the page. Exercises throughout ask you to modify the code, and each has a worked solution folded underneath it. Open the solution after you have tried; the failures are where the learning is.

Tip

The single best way to use this book is to work through it with your own dataset alongside the example. It is harder, and it is the difference between recognizing these concepts and owning them.

What to expect

The book has a plot, and it is not a flattering one. Chapter by chapter, results that look impressive turn out to be artefacts, and the methods that fix them introduce subtler problems of their own. Two villains recur: overfitting, which makes models look good on the data they were built from, and leakage, which makes them look good even on data they were not.

This is not pessimism. The point is that predictive modelling comes with a small number of specific failure modes, that they are well understood, and that avoiding them is mostly a matter of discipline rather than mathematics. By chapter 5 you will have a complete recipe for building a model and reporting its performance honestly.

The book focuses on simple, explainable models. But it also shows how the predictive framework lets you use a genuine “black box” without giving up confidence in what it does — because in this framework, trust comes from how a model was validated, not from whether you can read its coefficients.

A living document

This is not a conventional textbook. It is a Jupyter Book under continuous development. If something is wrong, unclear, or missing, please open an issue — that is how it gets better.

Have fun discovering a new way of dealing with your data.