Explaining Medical Model Exceptions - Computational Intelligence in Healthcare: Advanced Methodologies - page 254

Chapter 12
Explaining Medical Model Exceptions
Rainer Schmidt and Olga Vorobieva
Institute for Medical Informatics and Biometry, University of Rostock, Germany
[e-mail]
Abstract. In this chapter, a system named ISOR is presented, that supports re-
search doctors to investigate and to explain cases that do not fit a theoretical
hypothesis. The system is designed for situations where neither a well-
developed theory nor reliable knowledge nor, at the beginning, a case base is
available. Instead of theoretical knowledge and intelligent experience, just a
theoretical hypothesis and a set of measurements are given. ISOR is a Case-
Based Reasoning system. That means, when attempting to find an explanation
for an exceptional case, solutions of already explained similar exceptional cases
are considered. However, ISOR uses further knowledge sources, especially a
dialog where the user (a research doctor) can make suggestions for an explana-
tion. ISOR is domain independent and can be applied to various research prob-
lems. However, in this chapter, it is focused on the hypothesis that a specific
exercise program improves the physical condition of dialysis patients. Since
many data are missing for this research problem, a method to impute missing
data was developed and is also presented here. This method combines general
domain independent techniques with domain knowledge provided by a medical
expert. For the latter technique Case-based Reasoning is applied again.
1 Introduction
Case-based Reasoning (CBR) uses previous experience represented as cases to under-
stand and solve new problems. A case-based reasoner remembers former cases similar
to the current problem and attempts to modify solutions of former cases to fit the cur-
rent problem.
The fundamental ideas of CBR originated in the late eighties [1]. In the early nine-
ties CBR emerged as a method that was firstly described by Kolodner [2]. Later on
Aamodt and Plaza presented a more formal characterisation of the CBR method [3],
which consists of four steps: retrieving former similar cases, adapting their solutions
to the current problem, revising a proposed solution, and retaining new learned cases.
However, there are two main subtasks in Case-based Reasoning [2, 3]: The retrieval,
the search for a similar case, and the adaptation, the modification of solutions of re-
trieved cases. For retrieval a lot similarity measures and sophisticated retrieval algo-
rithms, have been developed within the CBR community. The most common ones are
indexing methods [2] like tree-hash retrieval [4], which are useful for nominal