Practical probabilistic inference

Abstract

The design and use of expert systems for medical diagnosis remains an attractive goal. One such system, the Quick Medical Reference, Decision Theoretic (QMR-DT), is based on a Bayesian network. This very large-scale network models the appearance and manifestation of disease and has approximately 600 unobservable nodes and 4000 observable nodes that represent, respectively, the presence and measurable manifestation of disease in a patient. Exact inference of posterior distributions over the disease nodes is extremely intractable using generic algorithms. Inference can be made much more efficient by exploiting the QMR-DT's unique structure. Indeed, tailor-made inference algorithms for the QMR-DT efficiently generate exact disease posterior marginals for some diagnostic problems and accurate approximate posteriors for others. In this thesis, I identify a risk with using the QMR-DT disease posteriors for medical diagnosis. Specifically, I show that patients and physicians conspire to preferentially report findings that suggest the presence of disease. Because the QMR-DT does not contain an explicit model of this reporting bias, its disease posteriors may not be useful for diagnosis. Correcting these posteriors requires augmenting the QMR-DT with additional variables and dependencies that model the diagnostic procedure. I introduce the diagnostic QMR-DT (dQMR-DT), a Bayesian network containing both the QMR-DT and a simple model of the diagnostic procedure. Using diagnostic problems sampled from the dQMR-DT, I show the danger of doing diagnosis using disease posteriors from the unaugmented QMR-DT.

(cont.) I introduce a new class of approximate inference methods, based on feed-forward neural networks, for both the QMR-DT and the dQMR-DT. I show that these methods, recognition models, generate accurate approximate posteriors on the QMR-DT, on the dQMR-DT, and on a version of the dQMR-DT specified only indirectly through a set of presolved diagnostic problems.