Abductive inference of events: diagnosing cardiac arrhythmias
Margaret E. Guertin · 1996
The ability to reason abductively, to infer a set of causes from their observed effects, is an essential component of human intelligence, and is used to construct explanations in such diverse areas as archaeology, criminal investigation, and medicine. Automating this process is, therefore, an important problem in artificial intelligence, but also an extremely challenging one. Its main difficulty lies in the exponential number of explanations for even the simplest set of observations. Abductive inference for real-world problems has been shown to be NP-hard. We offer a new approach to this problem--that of controlling the combinatoric explosion of possible explanations by means of temporal constraint propagation. To show the effectiveness of this approach, we have chosen to generate causal explanations of a cardiac arrhythmia appearing on an electrocardiogram (ECG) Explanations in this domain are constrained by well-understood temporal constraints from the ECG, as well as more subtle ones imposed by the underlying model of the heart. We demonstrate this approach with H scOLMES, an abductive reasoning system augmented by a method for propagating temporal constraints, which yields early refutation of inconsistent hypotheses. While H scOLMES is designed as a general-purpose system, it is used here to generate causes of cardiac arrhythmia. Though other arrhythmia diagnostic systems discover more arrhythmias than H scOLMES, they do so at a cost, in that they incorporate huge amounts of domain knowledge into their code, requiring a considerable investment of time and expertise. H scOLMES, on the other hand, provides very detailed explanations from a network of components embodying sparse amounts of domain knowledge and constrained by a weak set of temporal constraints. While H scOLMES generates more detailed explanations than are achieved in comparable diagnostic systems, its principal contribution lies in the alternative it offers to the traditional expert system/knowledge engineering approach to diagnostic problems. Many problems formerly thought to require immense amounts of domain-specific knowledge can, in fact, be solved efficiently by loosely connected networks of not very knowledgeable agents, provided that they are grounded in a good causal model of the problem.