Understanding Asthma

Globally around 450,000 people die each year from asthma. If we could better understand what causes people to develop asthma, it would have a hugely beneficial impact on asthma detection, diagnosis and treatment. Can model-based machine learning help provide this deeper understanding?

Asthma is a very common disease which affects around 5% of people in the UK [Anderson et al., 2007] and about 7% in the US [Fanta, 2009]. Asthma can have extremely serious outcomes for those who suffer from it. One known risk factor for developing asthma is if a person has allergies, but the relationship between allergies and asthma is not well understood. An improved understanding of this relationship could potentially allow early detection of the kind of severe asthma that can lead to hospitalisation or worse.

The Manchester Asthma and Allergy Study (MAAS) is a study designed to help understand the causes of childhood asthma and allergies [Custovic et al., 2002]. In particular, the study aims to understand why some children with allergies develop asthma while others do not. MAAS is a birth cohort study – in other words, people were recruited into the study at birth – and consists of around 1,000 people. The study began in 1995 and continues to this day, collecting ongoing data about the study participants, who are now young adults. As you might imagine, a huge amount of dedication and commitment is required of these participants and their families – we and the study team are immensely grateful to them all!

In this chapter, we will look at how to apply model-based machine learning to data collected in this study, to model the onset of childhood allergies and see how this relates to the development of asthma. This kind of machine learning application is different to those we have looked at in previous chapters, because we are interested in improving understanding as a primary goal of the project, rather than predicting who will develop asthma without any understanding of why. It’s worth looking at these two contrasting goals in a bit more detail:

• Predictive machine learning – the goal is to make predictions, without requiring an explanation of the predictions. This kind of goal is common when building automated systems where explanations are not needed.
• Explanatory machine learning – the goal is to explain or understand patterns in the data. This kind of goal is common when doing scientific or medical research, where there is a human in the loop who wishes to understand the processes that give rise to the data.

Often there are elements of both of these goals in a particular machine learning project. For example, when doing predictions it may be useful to provide some explanation of those predictions. And even when the primary goal is improved understanding, such as in this asthma project, it may still be useful to apply that understanding to make predictions, such as predicting whether a child will develop asthma.

The model developed in this chapter was created as part of a collaboration with the MAAS team, particularly Professors Adnan Custovic and Angela Simpson, as described in Simpson et al. [2010] and Lazic et al. [2013].

You can create results like those in this chapter using the companion source code [Diethe et al., 2019]. Since we cannot distribute the actual medical data used in this work, we have provided a synthetic data set that gives similar results to the true data.