Using Model-Based Machine Learning to Understand Childhood Asthma
Asthma and allergies rank among the most common chronic disorders in children—and their incidence is on the rise. The question is why. If we understood the underlying causes of asthma, we might learn how to prevent it in susceptible children and better treat it in adults.
Although evidence from twin studies suggests a strong genetic component in asthma and allergies, few of these studies have identified the same genetic associations. Moreover, the role of the environment in asthma and allergies is evidenced by the rapid increase in the prevalence of these disorders over the last four to five decades, a time frame too short to be attributable to genetic factors alone. And indeed, various environmental exposures have been associated with the development of asthma and allergies. However, as with genetics, the data on the role of environmental factors are inconsistent, with the same environmental exposure showing increased risk, protection, or no effect, depending on the study.
The conflicting evidence on the effects of genetic variants and environmental exposures may be due in part to these factors having largely been studied separately. By contrast, our research seeks to model genetic and environmental factors jointly. We view asthma as a complex disease that takes multiple forms, and therefore a central goal is to discover these underlying “phenotypes.” To do so, we are using a Bayesian model-based machine-learning approach, which, unlike conventional statistical analyses and black-box machine learning methods, easily allows the incorporation of rich, hierarchical structure derived from clinical background knowledge.
The project is a collaboration between Microsoft Research and the University of Manchester. Microsoft Research provides the machine-learning expertise through its team of Chris Bishop, John Winn, Markus Svensen, and Nevena Lazic in the Machine Learning and Perception group, and has contributed Infer.NET, a new framework that allows rapid construction of complex Bayesian models and performs efficient inference within those models. The University of Manchester brings world-class clinical expertise through Iain Buchan, Adnan Custovic, and Angela Simpson, along with a high-quality data set collected by the Manchester Asthma and Allergies Study.
Together, we hope to build complex models that represent a broad range of important variables associated with asthma. And while the immediate goal is to study the development of asthma, a successful outcome can highlight the benefits of a model-based approach to the analysis of clinical data generally, which could have much broader applicability.
This work significantly contributed to the development of Infer.NET by providing real-world testing of its scalability and capabilities, and led to the first paper published using this technology. Improvements to Infer.NET resulting from this project have helped prepare the framework to enhance a range of Microsoft products.
Learn more about this research: