Guide · Scientific writing

How to write a scientific paper abstract

Based on episode 278 of the podcast.

The abstract is the most-read part of your paper and the worst written. The editor-in-chief, the first gatekeeper, largely decides from it whether your work moves on. And 95% are written the same way: at the last minute, diving straight into the technical problem, heavy on methods, light on relevance. Only subfield experts understand them, and sometimes not even them.

After publishing more than 200 papers and reviewing many more, here is the method I use and teach in mentoring. It is not entirely mine: it builds on work where some authors deconstructed dozens of successful abstracts from journals like Nature and extracted their rules. I have added my own interpretation and a couple of tweaks.

This is the essentials, the framework I use, not a full treatise. Just enough to start well without wasting time. The detail of your specific case, we cover in 1:1 mentoring.

The core idea: the microscope

A good abstract goes from the most general to the most specific, then back to the general. Like a microscope: you start at low magnification (anyone understands), zoom to maximum detail (your result), then zoom back out to show why it matters. The opposite of the typical boring abstract, which dives into technical detail and never leaves.

The six-part method

All of this goes in one compact block of text. The “parts” are mental, not visual.

1. Very general introduction to the field. One or two sentences anyone on the street would understand.
2. Introduction for your field. One notch more specific: you speak to people in your discipline. More concrete, still not the problem yet.
3. The problem, in a single sentence. The open question in the field, phrased so even an outsider gets it. Keep the suspense.
4. Your main result, in one sentence, with impact. What you found, as clear and general as possible.
5. The comparison: why yours is different and important. Versus what was already known, what changes. This is your justification.
6. The general context: the implications. Close by zooming back out: what your work opens the door to.

The original work suggests a seventh “even broader context” part. I drop it: in my experience it confuses more than it helps.

Why it works

• It impacts the editor-in-chief, who is almost never from your exact subfield: they grasp your contribution without being a specialist.
• It raises your odds of passing the first filter.
• It is reusable: the same text, adapted, works for the grant application, the talk or the social thread.

Worked example

First, how this abstract would look written badly, like the 95%, and then the one I actually published.

Before (weak version, dives straight into methods and never leaves):

In this work, artificial neural networks (ANN), logistic regression (LR) and support vector machines (SVM) were trained on several medical datasets and evaluated through cross-validation. Hyperparameters were tuned via grid search and performance was measured using accuracy, F1-score and AUC. Convolutional neural networks (CNN) were additionally implemented on GPU and HPC infrastructure for image data. The models achieved competitive accuracy across the tested datasets, and feature-importance values were computed for some of them. Future work will explore further architectures and datasets.

Only four specialists understand it, it never states the problem, compares to nothing and never explains why a doctor would care. Now the real one, my survey Towards the Interpretability of Machine Learning Predictions for Medical Applications Targeting Personalised Therapies: A Cancer Case Survey (Int. J. Mol. Sci., 2021):

After (the published one, with the six parts marked):

Artificial Intelligence is providing astonishing results, with medicine being one of its favourite playgrounds (1). Machine Learning and, in particular, Deep Neural Networks are behind this revolution. Among the most challenging targets of interest in medicine are cancer diagnosis and therapies, but to start this revolution, software tools need to be adapted to cover the new requirements (2). In this sense, learning tools are becoming a commodity but, to be able to assist doctors on a daily basis, it is essential to fully understand how models can be interpreted (3). In this survey, we analyse current machine learning models and other in-silico tools as applied to medicine, specifically, to cancer research, and we discuss their interpretability, performance and the input data they are fed with (4). Artificial neural networks, logistic regression and support vector machines have been observed to be the preferred models; convolutional neural networks, supported by GPUs and HPC, are gaining importance for image processing (5). However, the interpretability of machine learning predictions, so that doctors can understand them, trust them and gain useful insights for clinical practice, is still rarely considered, a factor that needs improving to enhance doctors’ predictive capacity and achieve individualised therapies in the near future (6).

Notice: it opens with a sentence anyone understands (1), drops into field detail (2,3), states what it does (4), gives the findings (5) and closes back at the broad level, personalised therapies, (6). The whole microscope: general → specific → general.

The mistakes I see again and again

• Writing the abstract on the last day, any old way.
• Starting with the technical problem only four people understand.
• Three sentences of methodology, none on relevance.
• Not comparing your result with what came before.
• Ending without saying why it matters.

And the title

The same principle applies to the title: impact before exhaustiveness. The survey’s title, “Towards the Interpretability of Machine Learning Predictions for Medical Applications Targeting Personalised Therapies: A Cancer Case Survey”, tells you exactly what you’ll find: interpretability, medicine, personalised therapies, cancer. It communicates the what and the why, not the machinery.