" In this article I share with the readers the basics of and the principles behind intelligent decision support systems based on the theoretically sound principles of probability theory and decision theory. Recent advances in probability theory have led to the development of graphical models that are capable of modeling the causal structure of systems, well understood by human experts, and at the same time give such structure a sound, probabilistic interpretation. Graphical models can serve as a convenient basis for modeling domains that involve high degree of uncertainty and also reasoning with them. Examples of problems that are addressed by systems based on graphical models include computer vision, robotics, pattern matching, medical diagnosis and therapy planning, machine diagnosis, and even on-line help. Microsoft Corporation uses graphical models inside the Windows operating system, in troubleshooting, and in user interfaces, such as on-line Office help and junk mail filtering in Outlook. The graphical models methodology is implemented in a general purpose decision modeling system SMILE and its Windows user interface, GeNIe, developed at the Decision Systems Laboratory. I try to give the readers a flavor of GeNIe models and building systems based on the SMILE library. "

("One-page extended abstract of the conference presentation of GeNIe and SMILE").

(Two-page extended abstract of conference presentation of GeNIe and SMILE.)

" Constructing diagnostic Bayesian network models is a complex and time consuming task. In this paper, we propose a methodology to simplify and speed up the design of such models. The models are based on two simplifying assumptions: (1) the structure of the model has three levels of variables and (2) the interaction among the variables can be modelled by noisy-MAX gates. The methodology is implemented in an application named: GeNIeRate, which aims at supporting construction of diagnostic Bayesian network models consisting of hundreds or even thousands of variables. Preliminary qualitative evaluation of this application shows great promise. We are planning to conduct a systematic study to compare GeNIeRate to traditional techniques for building Bayesian network models and we hope to be able to present the results at the workshop. "

" Building probabilistic and decision-theoretic models requires a considerable knowledge engineering effort in which the most daunting task is obtaining the numerical parameters. Authors of Bayesian networks usually combine various sources of information, such as textbooks, statistical reports, databases, and expert judgement. In this paper, we demonstrate the risks of such a combination, even when this knowledge encompasses such seemingly population-independent characteristics as sensitivity and specificity of medical symptoms. We show that the criteria "do not combine knowledge from different sources" or "use only data from the setting in which the model will be used" are neither necessary nor sufficient to guarantee the correctness of the model. Instead, we offer graphical criteria for determining when knowledge from different sources can be safely combined into the general population model. We also offer a method for building subpopulation models. The analysis performed in this paper and the criteria we propose may be useful in such fields as knowledge engineering, epidemiology, machine learning, and statistical meta-analysis. "

" We describe our work on HEPAR II, a probabilistic causal model for diagnosis of liver disorders. The model, a Bayesian network capturing the causal interactions among various risk factors, diseases, symptoms, and test results, is based on expert knowledge combined with clinical data captured in medical records. The main applications of HEPAR II are assistance is diagnosis and training of beginning diagnosticians. We outline the principles of the applied approach, present a brief description of the model, and report its diagnostic performance. "

" Missing values of attributes in data sets, also referred to as incomplete data, pose difficulties in learning tasks, such as classification, data mining, or learning Bayesian network structure and its numerical parameters. Because of the predominance of incomplete data in practice, many methods have been proposed to deal with them while there are few studies that compare their performance. The HEPAR II project presents an excellent opportunity to test experimentally how these methods perform on a real data set. We briefly review several popular methods for handling incomplete data and then compare them on the task of learning conditional probability distributions of a Bayesian network model, where the comparison criterion is the resulting diagnostic accuracy. While substitution of "normal" values of missing attributes seemed to perform best, we observed only a small difference in performance among the studied methods. "

" While most knowledge engineers believe that the quality of results obtained from Bayesian networks is not too sensitive to imprecision in probabilities, this remains a conjecture with only modest empirical support. Our work on a Bayesian network model for diagnosis of liver disorders, Hepar II, presented us with an opportunity to test this conjecture in a practical setting. We present the results of an empirical study in which we systematically introduce noise in Hepar II's probabilities and test the diagnostic accuracy of the resulting model. We replicate an experiment conducted by Pradhan et al. [13] and show that Hepar II is more sensitive to noise in parameters than the CPCS network that they examined. Our data show that the diagnostic accuracy of the model deteriorates almost linearly with noise. While our result is merely a single data point that sheds light on the hypothesis in question, we suggest that Bayesian networks are more sensitive to the quality of their numerical parameters than popularly believed. "

" Almost two decades after the introduction of probabilistic expert systems, their theoretical status, practical use, and experiences are matching those of rule-based expert systems. Since both types of systems are in wide use, it is more than ever important to understand their advantages and drawbacks. We describe a study in which we compare rule-based systems to systems based on Bayesian networks. We present two expert systems for diagnosis of liver disorders that served as the inspiration and vehicle of our study and discuss problems related to knowledge engineering using the two approaches. We finally present the results of a simple experiment comparing the diagnostic performance of each of the systems on a subset of their domain. "

" The last two decades have brought considerable advances in the field of computer-based medical systems. These advances have resulted in noticeable improvements in medical care, starting from ease of storage and access of digital imaging through gathering of computerized medical data, accessing on-line literature, patient monitoring, and therapy planning. Systems addressing the task of diagnosis, however, have rarely been adopted in clinical practice, which has raised questions about their usefulness and feasibility. We show in this paper a practical application of a computer-based diagnostic system, HEPAR II, to training beginning diagnosticians. "

" This work proposes a model-based guideline development method for complex diagnostic and treatment policies, for instance for symptom-based indications.
An influence diagram has advantages for the development of outcome-based guidelines, because it provides an explicit link between the policies and these outcomes. A Bayesian network have a strong inference capability, suitable for complex symptom-based medical diagnostic modelling incorporating many diseases. However, neither supports temporal sequences.
The modelling of symptom-based indications for guideline distillation can be supported by imposing a strict structure and a strict order. A further extension is the user interface for evaluation and ranking of the probabilities, policies, and the resulting outcomes. Instead of seeking for the best solution, the guideline is distilled by ranking these outcomes with respect to the treatment, anatomic region and pathology, and the generally expected outcome if treated properly. The proposed guideline distillation method is expected to improve the development, and therefore the quality of the guidelines. "