Case-based prediction in experimental medical studies

https://doi.org/10.1016/S0933-3657(98)00057-8Get rights and content

Abstract

Case-based approaches predict the behaviour of dynamic systems by analysing a given experimental setting in the context of others. To select similar cases and to control adaptation of cases, they employ general knowledge. If that is neither available nor inductively derivable, the knowledge implicit in cases can be utilized for a case-based ranking and adaptation of similar cases. We introduce the system OASES and its application to medical experimental studies to demonstrate this approach.

Introduction

Simulation is an experiment performed on a model that is aimed at answering questions about the behaviour of a system. The experiment we are interested in is analysing and predicting the effect of settings in experimental medical studies. The subjects of these experimental studies define the system, namely the organisms that the model is intended to represent. Modelling means the process of organising knowledge about a given system [26]. Thus the model design determines the way how knowledge is structured and how easily knowledge can be acquired.

Typically, simulation employs an explicit model of the objects, interrelations and causal forces of the domain, as many diagnostic systems do [16], [14]. The model might comprise quantitatively and qualitatively scaled variables to describe the system and its components. In many areas, prediction is hampered by an incomplete domain model and the difficulty in defining general rules. Purely qualitative deductive modelling and simulation provide no satisfying answer to this problem. Deductive qualitative approaches are designed to deal with a set of variables, whose values are not totally determined. However, they require a well defined and rather simple structure [5]. If sufficient and uniform data are available, inductive methods can be applied in order to derive general rules of behaviour and to complete the domain model [17]. The method of fuzzy-based inductive reasoning (FIR) has demonstrated how the accuracy of purely inductive forecasting is increased by integrating and using cases for forecasting [20]. Not only in the context of forecasting, case-based approaches have demonstrated their strength compared to other approaches like model-based, purely inductive or rule-based systems [15], [4]. Case-based reasoning also proved to be useful in diagnosis and planning tasks [1], [18]. In FIR, experimentation is based on a model that is implicitly encoded in a set of cases, a similarity measure, and an adaptation procedure. The retrieval of cases is guided by inductively derived general knowledge and requires time series, i.e. huge amounts of homogeneously structured data [24]. This approach is designed for the work on first order data, but is doomed to fail on second-order data, which summarize the results of experimental studies. Typically, each experimental study is conceptualized and processed independently of others or even with the intention to deviate from existing ones. Therefore the structures of the studies often differ substantially.

The objective of our research is the development of a system that supports the experimentation with a model that is based on second order data, namely the results of experimental studies typically found in the literature. Like other case-based approaches, the model for experimentation is given in terms of cases. However, neither guidance of retrieval by induction nor usage of general knowledge for the definition of similarity and adaptation is possible. Similarity and adaptation of cases are rather case sensitive, i.e. they are dependent on the context. Differences can only be interpreted with respect to the case. In the same way, adaptation procedures are based on the special features of the adapted case. To answer what results are to be expected in given experimental settings, the system OASES, our approach to simulation based on experimental studies, intensively uses the knowledge inherent in cases.

OASES has first been tested in the area of bone healing, where the general significance and importance of single factors is still unknown. Factors that have an effect on fracture healing are both local, for example the degree of local trauma, vascular injury or infection, and systemic, like age or hormones. Besides biological interventions like osteogenic, osteoconductive and osteoinductive methods, mechanical and physical interventions and their influence on bone formation are analyzed [10]. The large number of factors that are influencing fracture healing is found in many different experimental investigations, which are often based on different species. Sometimes the data are described quantitatively, but often only qualitatively. This diversity hampers the comparison of experimental studies as it hinders the development and validation of a deductive model that would help to assess the influence of a given experimental setting on the healing process.

Often comparative statements can be found in experimental studies, as those typically analyse the effect of single factors in a given experimental setting. For example, ‘more proliferative callus formation was seen in the less-rigidly fixed groups’ ([12], p. 38). This salient feature of experimental studies is found in the area of bone healing and constitutes the backbone of our approach.

Section snippets

The case base

The case base is the only source of knowledge the system uses. Therefore, the results of experimental studies have to be encoded in a set of cases. In the domain of bone healing experimental studies analyse the influence of factors and interventions on the healing process. The experimental setting can be expressed by a set of experimental parameters, for example ‘species’ or ‘osteotomy’, whereas the results are measured by result parameters, for example the ‘torsional stiffness’ of the fixed

Evaluation

OASES was evaluated based on 650 cases, which condense experimental results that are documented in around 50 publications in the domain of bone healing. For evaluation, one selected publication was removed from the case base and used for testing. About 90 inquiries on the most frequently used experimental parameters were made by this means and compared with the de facto results of the removed publication. Below we present some of the results produced by OASES.

In Claes, Wilke et al., 1989 [7],

Conclusion

OASES supports prediction and experimentation based on models that are encoded in cases. Sets of cases represent experimental studies. The system consequently employs the ideas of case-based reasoning throughout the whole process of prediction. The heterogeneity of second order data, which can be found in experimental studies on bone healing, requires specific context sensitive methods for matching and adaptation. Usually, experimental studies exactly provide knowledge for that methods, i.e.

Acknowledgements

The research was sponsored by the Forschungsschwerpunktprogramm Baden Wuerttemberg. We would like to thank the Department of Orthopaedic Research and Biomechanics at the University of Ulm for providing us with the necessary literature, and particularly Steffen Wolf for his constructive criticism and discussions, which helped us to develop the system.

References (26)

  • R. Duda et al.

    Pattern Classification and Scene Analysis

    (1973)
  • T.A. Einhorn

    Enhancement of fracture-healing

    J. Bone. Joint. Surg. Am.

    (1995)
  • A.E. Goodship et al.

    The role of fixator frame stiffness in the control of fracture healing; an experimental study

    J. Biomech.

    (1993)
  • Cited by (0)

    View full text