Maximal prime subgraph decomposition of Bayesian networks.

The authors present a method for decomposition of Bayesian networks into their maximal prime subgraphs. The correctness of the method is proven and results relating the maximal prime subgraph decomposition (MPD) to the maximal complete subgraphs of the moral graph of the original Bayesian network are presented. The maximal prime subgraphs of a Bayesian network can be organized as a tree which can be used as the computational structure for LAZY propagation. We also identify a number of tasks performed on Bayesian networks that can benefit from MPD. These tasks are: divide and conquer triangulation, hybrid propagation algorithms combining exact and approximative inference techniques, and incremental construction of junction trees. We compare the proposed algorithm with standard algorithms for decomposition of undirected graphs into their maximal prime subgraphs. The discussion shows that the proposed algorithm is simpler, more easy to comprehend, and it has the same complexity as the standard algorithms.

Practical modeling of Bayesian decision problems - exploiting deterministic relations.

The widespread use of influence diagrams to represent and solve Bayesian decision problems is still limited by the inflexibility and rather restrictive semantics of influence diagrams. We propose a number of extensions and adjustments to the definition of influence diagrams in order to make the practical use of influence diagrams more flexible and less restrictive. In particular, we describe how deterministic relations can be exploited to increase the flexibility and efficiency of representing and solving Bayesian decision problems. The issues addressed in the paper were motivated by the construction of a decision support system for mission management of unmanned underwater vehicles (UUVs).

A method for diagnosing multiple diseases in MUNIN.

A new method for diagnosing multiple diseases in large medical decision support systems based on causal probabilistic networks is proposed. The method is based on characteristics of the diagnostic process that we believe to be present in many diagnostic tasks, both inside and outside medicine. The diagnosis must often be made under uncertainty, choosing between diagnoses that each have small prior probabilities, but not so small that the possibility of two or more simultaneous diseases can be ignored. Often a symptom can be caused by several diseases and the presence of several diseases tend to aggravate the symptoms. For diagnostic problems that share these characteristic, we have proposed a method that operates in a number of phases: in the first phase only single diseases are considered and this helps to focus the attention on a smaller number of plausible diseases. In the second phase, pairs of diseases are considered, which make it possible to narrow down the field of plausible diagnoses further. In the following phases, larger subsets of diseases are considered. The method was applied to the diagnosis of neuromuscular disorders, using previous experience with the so-called MUNIN system as a starting point. The results showed that the method gave large reductions in computation time without compromising the computational accuracy in any substantial way. It is concluded that the method enables practical inference in large medical expert systems based on causal probabilistic networks.

The EMG diagnosis--an interpretation based on partial information.

There is a large difference between the prevalence of a given disease in the general population and in the population seen in the EMG lab. It can be argued that both prevalences are the correct choice as prior probabilities for the diseases. This paradox is resolved by recognizing that the EMG diagnosis is only based on the information provided by the EMG examination and thus only represents a partial view of the patient. We propose a solution summarizing the set of findings, signs and symptoms, lab results etc., that led to the referral of the patient for an EMG examination. This information is described by stochastic variables called FIDL factors (Found In Doctor's Lab). The approach is tested on the EMG expert system MUNIN with 30 previously evaluated cases. The results show that this solution improves the specificity of the diagnosis, without affecting the sensitivity.

Evaluation of the diagnostic performance of the expert EMG assistant MUNIN.

The diagnostic performance of the medical expert system MUNIN for diagnosis of neuromuscular disorders was evaluated on a set of 30 test cases. The cases were provided by 7 experienced electromyographers who were subsequently invited to participate in the evaluation. To reasonably cover the range of disorders, the electromyographers were asked to provide cases from patients with different types of muscular dystrophy, with neuromuscular transmission disorders, with motor neurone disease, and with different types of polyneuropathies. In addition, patients with a range of local neuropathies were provided. Out of the 30 cases, 11 cases were evaluated by an "almost peer review" method and the remaining 19 cases were evaluated by a "silver standard" method. The number of cases evaluated by "almost peer review" was limited to 11 due to time constraints on the evaluation procedure. During the "almost peer review," each electromyographer was asked to diagnose patients, using a vocabulary that closely resembled MUNIN's vocabulary. Subsequently, we attempted to provide a consensus diagnosis for the patients based on discussion among the participating electromyographers. The electromyographers were also asked to assess how well MUNIN had performed in each case. The remaining 19 cases were evaluated by a "silver standard" procedure, where MUNIN's diagnosis was compared to the diagnosis of the expert who provided the case. The results indicated that MUNIN performed well, and the electromyographers considered "that MUNIN performed at the same level as an experienced neurophysiologist." In particular, it was noted that MUNIN handled cases with conflicting findings well, and that it was able to diagnose patients with multiple diseases.

A probabilistic approach to glucose prediction and insulin dose adjustment: description of metabolic model and pilot evaluation study.

A model of carbohydrate metabolism has been implemented as a causal probabilistic network, allowing explicit representation of the uncertainties involved in the prediction of 24-h blood glucose profiles in insulin-dependent diabetic subjects. The parameters of the model were based on experimental data from the literature describing insulin and carbohydrate absorption, renal loss of glucose, insulin-independent glucose utilisation and insulin-dependent glucose utilisation and production. The model can be adapted to the observed glucose metabolism in the individual patient and can be used to generate predicted 24-h blood glucose profiles. A penalty is assigned to each level of blood glucose, to indicate that high and low blood glucose levels are undesirable. The system can be asked to find the insulin doses that result in the most desirable 24-h blood glucose profile. In a series of 12 patients, the system predicted blood glucose with a mean error of 3.3 mmol/l. The insulin doses suggested by the system seemed reasonable and in several cases seemed more appropriate than the doses actually administered to the patients.

Specification of models in large expert systems based on causal probabilistic networks.

Problems involved in the specification of large expert systems are discussed. In the specification of causal probabilistic networks conditional probability tables for all nodes have to be provided. These conditional probability tables can often be described by models that specify the nature of interaction between nodes. Various types of models are described and a program that handles such models is presented. Large causal probabilistic networks often contain several copies of identical tables or structures. A header facility that provides common definitions of such repeated elements is proposed. This facility makes specifications much shorter and easier to construct and maintain.

Diagnostic function of the microhuman prototype of the expert system--MUNIN.

This paper describes the diagnostic function of a prototype expert system for electromyography (EMG). The prototype was restricted to a limited "Microhuman" anatomy with only 6 muscles and 8 nerves, and a corresponding limitation on the number of local nerve lesions. It attempted to give a detailed description of the most important groups of generalized nerve and muscle disorders, and the commonly used parameters from needle EMG and nerve conduction studies were included. The system can be used both for "diagnostic" and for "causal" reasoning. In diagnostic reasoning, the system's probabilistic inference engine is used to reason from test results through 14 different aspects of neuromuscular pathophysiology to disorders. In causal reasoning, the system reasons in the opposite direction from disorders through pathophysiology to expected test results. The diagnostic function of the system was illustrated by 3 cases: a normal subject, a patient with a bilateral carpal tunnel syndrome and a patient with both a diabetic polyneuropathy and a bilateral carpal tunnel syndrome.

Medical expert systems based on causal probabilistic networks.

Causal probabilistic networks (CPNs) offer new methods by which you can build medical expert systems that can handle all types of medical reasoning within a uniform conceptual framework. Based on the experience from a commercially available system and a couple of large prototype systems, it appears that CPNs are now an attractive alternative to other methods. A CPN is an intensional model of a domain, and it is therefore conceptually much closer to qualitative reasoning systems and to simulation systems than to rule-based or logic-based systems. Recent progress in Bayesian inference in networks has yielded computationally efficient methods. The inference method used follows the fundamental axioms of probability theory, and gives a sound framework for causal and diagnostic (deductive and abductive) reasoning under uncertainty. Experience with the prototypes indicates that it may be possible to use decision theory as a rational approach to test planning and therapy planning. The way in which knowledge is acquired and represented in CPNs makes it easy to express 'deep knowledge' for example in the form of physiological models, and the facilities for learning make it possible to make a smooth transition from expert opinion to statistics based on empirical data.