A major part of the medical knowledge concerns diseases that are uncommon or even rare. The uncommon nature of these disorders renders it impossible to collect data of a sufficiently large number of patients to develop machine-learning models that faithfully reflect the subtleties of the domain. An alternative is to develop a Bayesian network with the help of clinical experts. Lack of data is then compensated for by eliciting the structure with its associated local probability distributions from the experts. The resulting network can be subsequently evaluated using the available dataset. One may also consider adopting very strong independence assumptions, such as in naive Bayesian models. Normally not all subtleties of the interactions among the variables in the domain are reflected in such models. Yet, a relatively small dataset suffices to obtain an acceptably accurate model. This paper explores the trade-offs between modelling using expert knowledge, and machine learning using a small clinical dataset in the context of Bayesian networks.
展开▼