Expert systems are built from knowledge traditionally elicited from the human expert. It is precisely knowledge elicitation from the expert that is the bottleneck in expert system construction. On the other hand, a data mining system, which automatically extracts knowledge, needs expert guidance on the successive decisions to be made in each of the system phases. In this context, expert knowledge and data mining discovered knowledge can cooperate, maximizing their individual capabilities: data mining discovered knowledge can be used as a complementary source of knowledge for the expert system, whereas expert knowledge can be used to guide the data mining process. This article summarizes different examples of systems where there is cooperation between expert knowledge and data mining discovered knowledge and reports our experience of such cooperation gathered from a medical diagnosis project called Intelligent Interpretation of Isokinetics Data, which we developed. From that experience, a series of lessons were learned throughout project development. Some of these lessons are generally applicable and others pertain exclusively to certain project types.

Abstract

There are now domains where information is recorded over a period of time, leading to sequences of data known as time series. In many domains, like medicine, time series analysis requires to focus on certain regions of interest, known as events, rather than analyzing the whole time series. In this paper, we propose a framework for knowledge discovery in both one-dimensional and multidimensional time series containing events. We show how our approach can be used to classify medical time series by means of a process that identifies events in time series, generates time series reference models of representative events and compares two time series by analyzing the events they have in common. We have applied our framework on time series generated in the areas of electroencephalography (EEG) and stabilometry. Framework performance was evaluated in terms of classification accuracy, and the results confirmed that the proposed schema has potential for classifying EEG and stabilometric signals. The proposed framework is useful for discovering knowledge from medical time series containing events, such as stabilometric and electroencephalographic time series. These results would be equally applicable to other medical domains generating iconographic time series, such as, for example, electrocardiography (ECG).

SummaryObjectives: We present a framework specially designed to deal with structurally complex data, where all individuals have the same structure, as is the case in many medical domains. A structurally complex individual may be composed of any type of singlevalued or multivalued attributes, including time series, for example. These attributes are structured according to domain-dependent hierarchies. Our aim is to generate reference models of population groups. These models represent the population archetype and are very useful for supporting such important tasks as diagnosis, detecting fraud, analyzing patient evolution, identifying control groups, etc. Methods: We have developed a conceptual model to represent structurally complex data hierarchically. Additionally, we have devised a method that uses the similarity tree concept to measure how similar two structurally complex individuals are, plus an outlier detection and filtering method. These methods provide the groundwork for the method that we have designed for generating reference models of a set of structurally complex individuals. A key idea of this method is to use event-based analysis for modeling time series. Results: The proposed framework has been applied to the medical field of stabilometry. To validate the outlier detection method we used 142 individuals, and there was a match between the outlier ratings by the experts and by the system for 139 individuals (97.8%). To validate the reference model generation method, we applied k-fold cross validation (k = 5) with 60 athletes (basketball players and ice-skaters), and the system correctly classified 55 (91.7%). We then added 30 non-athletes as a control group, and the method output the correct result in a very high percentage of cases (96.6%). Conclusions: We have achieved very satisfactory results for the tests on data from such a complex domain as stabilometry and for the comparison of the reference model generation method with other methods. This supports the validity of this framework.

No abstract