New types of probabilistic graphical modelsapplications to medicine

Abstract

Probabilistic graphical models (PGMs) play a major role in much of the mod- ern research in reasoning with uncertainty, decision analysis, planning, pattern recognition, and many other areas. Several types of PGMs have been proposed in the last two decades. However, there are some problems for which none of these types are appropriate. For example, none of the types of PGMs proposed has been widely adopted for representing and solving asymmetric decision problems. Decision analysis networks (DANs) have been recently proposed by our research group and they needed efficient evaluation algorithms in order to be applicable to real-world problems. In this thesis, I propose a new algorithm that evaluates DANs by recursively decomposing them into a set of symmetric DANs, which can then be evaluated with standard algorithms, such as variable elimination or arc reversal. The efficiency of this algorithm matches that of the algorithms proposed for other asymmetric representations. Similarly, existing types of PGMs were not apt as dynamic modeling methods for cost-effectiveness analysis (CEA). The existing dynamic PGMs are burdened by the complexity of their evaluation and can only solve unicriterion problems. Only Markov infuence diagrams (MIDs), a more restricted type of dynamic PGMs also proposed by our research group, are suitable to build complex dy- namic models to perform CEA. I have developed new types of potentials and new sensitivity analysis algorithms, with which I have been able to replicate as MIDs several models proposed in the literature and to build two new models for CEA: one for malignant pleural effusion and another one for mammography screening. Finally, with the help of an expert, we have built a decision-support system for cochlear implant programming (i.e., parameter tuning) based on PGMs. In 1 this thesis, we also describe tuning networks, a new type of PGM we developed because existing PGMs were not suitable to model the behavior of systems with a high number of tunable parameters. This decision-support system is now routinely used at a hearing clinic in Antwerp (Belgium) to assist audiologists in the programming of cochlear implants. All the contributions to PGMs described in this thesis have been imple- mented in OpenMarkov, an open-source software tool developed at the UNED, and are publicly available.