A temporal representation and reasoning model for medical decision-support systems
Gregory F. Cooper, Constantin Aliferis · 1998
Reasoning with temporal concepts is an essential part of medical problem solving. As a result, medical decision-support systems need effective formalisms (i.e., representation and reasoning languages) for time modeling. Time modeling has been extensively studied in many disciplines. Consequently, we know today methods for handling important aspects of time modeling, as well as a number of unsolved conceptual and engineering challenges. On the basis of previous research, this thesis proposes a set of desiderata for formalisms for medical time modeling. The dissertation addresses 3 hypotheses: It is possible to construct a formalism that meets the time-modeling desiderata (hypothesis #1). It is possible to apply the new formalism to a medical domain of high temporal complexity and demonstrate expressivity (hypothesis #2) and tractability (hypothesis #3) advantages over alternative established formalisms. Current formalisms for temporal modeling achieve tractability by significantly constraining the temporal representations they use and the inferences they can make, or by abolishing formal temporal semantics. Since high expressiveness and explicitness are part of the time-modeling desiderata, this work adopts a different strategy, based on two conceptual principles: Hybrid temporal abstraction/explicitness allows the simultaneous representation of different levels of temporal abstraction and explicitness in the same model. This flexibility prevents unnecessary model size growth. Dynamic temporal abstraction constructs a small/explicit (and relatively more tractable) model that is pertinent to a query or patient from a large/implicit (and relatively less tractable) model that is pertinent to a domain or population. Using the above principles, Bayesian belief networks are extended to represent time and causality. The new formalism is called Modifiable Temporal Belief Networks (MTBNs). Formal definitions, properties, inference procedures and modeling techniques are given for MTBNs. It is shown that MTBNs address the desiderata for time modeling in medical decision-support systems. An initial formative study of MTBNs is conducted by applying them in the domain of liver transplantation. The liver modeling experiment suggests that MTBNs have expressivity and tractability advantages over alternative formalisms. This work concludes by discussing current weaknesses of MTBNs and corresponding research problems.