Integrating Deep Biomedical Models into Medical Decision Support Systems: An Interval Constraint Approach
Lecture Notes in Computer Science
Knowledge representation has always been a major problem in the design of medical decision support systems. In this paper we present a new methodology to represent and reason about medical knowledge, based on the declarative specification of interval constraints over the medical concepts. This allows the integration of deep medical models involving differential equations developed in biomedical research (typical in several medical domains) which, due to their complexity, have not been
... ed into medical decision support systems. The methodology which enables reasoning both forward and backward in time, is applied to a specific domain, electromyography. The promising results obtained are discussed to justify our future work. medical knowledge" ,  . In contrast with systems that represented causal but "shallow" relations, say between a disease and a symptom, they represent a more detailed set of pathophysiologic states and processes to explain that shallow relation. This approach has to handle certain difficulties. On the one hand, the network of concepts and relations is larger, imposing efficient methods to explore them. On the other hand, the deeper the knowledge is, the more dependent it is in numerical data. Reasoning with this kind of information is, in principle, very costly. Nevertheless, some formalisms may be quite efficient to handle it, namely belief networks  . Eliciting all the conditional probabilities in such a belief network is however a difficult problem, made only harder when several diseases are considered, due to the explosion of the required number of conditional probabilities . An alternative (or perhaps complementary) approach is to adopt "deep" biomedical models proposed and accepted in the medical community, but not incorporated into the decision support systems, for example, Cardiovascular , Respiratory  and Compartment models  . But these are often highly non-linear, based on differential equations, and difficult to reason about by simple "logical" procedures. Simulation of these models can be used to validate some hypotheses. However the usual simulation methods have difficulties to cope with uncertain data (e.g. they might provide "wrong" approximate solutions), and they are "time-directed": given the initial conditions the state of the system can be computed at some later time, but not at some previous time. As such they may only play a "passive" role in a task such as diagnosis, where the goal is to find a problem that has occurred in the past. In this paper we explore constraint solving over intervals, a recently proposed technology to handle non-linear numerical models, and its use in neurology. Next section presents the model we are using (based on differential equations). Section 3, addresses the advantages of constraint technology in the medical domain with some very simple examples. Section 4 discusses the handling of differential equations and highlights our contributions in this area. Section 5 shows preliminary results obtained with the use of this technology in modelling local nerve lesions. Finally, section 6 presents the major conclusions and discusses our plans for future work.