Our public health systems are faced with enormous opportunities: new possiblities for diagnosis and treatment, personalized medicine, the availability of large data sets. At the same time, health cost and the demand on the system is increasing at an unsustainable rate.
In this talk, I will illustrate how Bayesian machine learning can contribute to solve some of these problems, by discussing two examples: full Bayesian Gaussian Process.
Regression for learning non-linear interactions and finding missing heritability in genetic data; and a graphical model expert system that can assist medical specialists to diagnose patients using Bayesian inference.
For many traits and common human diseases, genetic association studies account for little of the known heritable variation. We suggest that this “Missing heritability” might lie in the effect of non-additive interactions between multiple loci. We employed a non-parametric, Bayesian method, based on Gaussian Process Regression. We analysed 46 quantitative yeast phenotypes and found that over 70% of the total known missing heritability could be explained, significantly improving on existing methods. Importantly, the availability of biological replicates significantly improved the power to identify such loci and, hence, to explain variance. These results represent a significant advance in approaches to understanding the missing heritability problem with potentially important implications for studies of complex, quantitative traits. Joint work with Kevin Sharp, Wim Wiegerinck, Alejandro Arias Vaquez, Barbara Franke and Kees Albers.
Diagnostic errors in US hospitals are estimated to occur in about 5 percent of adults patients, half of which are potentially harmful. In addition, (junior) specialists are known to request too many tests in order to arrive at a diagnosis, significantly increasing the cost of diagnosis. These facts motivate the need for a clinical expert system that assists doctors in the diagnostic process. Despite their obvious advantages, such systems have not been deployed in practice very often. We discuss some of the possible reasons. I will then show how to represents medical knowledge in terms of a Bayesian network. It computes a distribution over the most likely combination of diagnoses, and their probabilities using Monte Carlo sampling. I will illustrate the approach on some real patient cases. Joint work with Anna Simons, Radboud Medical Center Nijmegen.
Disclosure of Interest None declared