Using SMILE Wrappers

Diagnosis is one of the most successful applications of Bayesian networks. The ability of probabilistic knowledge representation techniques to perform both predictive and diagnostic inference makes them particularly suitable for diagnostic tasks. Bayesian networks can combine, or fuse, various sources of information—such as patient or equipment history, risk factors, symptoms, and test results—into a coherent probabilistic framework.

A Bayesian network built with SMILE can represent different components of a system, potential faults (or disorders), their observable manifestations (symptoms or alarms), and the results of diagnostic tests. In essence, such a model captures how possible defects—whether in a natural system such as the human body or in an engineered system such as a car, aircraft, or copier—may produce observable effects like error messages, abnormal readings, or test outcomes.

To perform diagnostic analysis, certain nodes in the network must be designated as diagnostic faults, while others must be marked as diagnostic observations. Optionally, observation nodes can also be assigned observation costs, representing, for example, the financial or time cost of performing a medical test or measurement.

The output of SMILE’s diagnostic algorithms is twofold:

1.Posterior marginal probabilities of diagnostic faults (reflecting their likelihoods given the available evidence).

2.A ranking of potential observations, ordered from most to least informative with respect to the current diagnostic focus.

Depending on the selected diagnostic algorithm, the observation ranking is based on one of the following measures:

•Cross-entropy – an information-theoretic, utility-free measure that estimates the expected reduction in entropy (uncertainty) over a specified subset of fault states after observing a particular node X. It quantifies the expected value of an observation for improving diagnostic certainty.

•Distance metrics – measures of difference between two probability vectors representing the pursued fault states before and after observation. SMILE supports several distance metrics, including Euclidean, cosine, and city-block (Manhattan) distances.

Tutorial 9 demonstrates diagnostic reasoning using the HeparII medical diagnosis model, illustrating both single- and multi-fault diagnosis scenarios.