Medical data mining using evolutionary computation
Introduction
Data mining aims at discovering novel, interesting and useful knowledge from databases [9]. Conventionally, the data is analyzed manually. Many hidden and potentially useful relationships may not be recognized by the analyst. Nowadays, many organizations including modern hospitals are capable of generating and collecting a huge amount of data. This explosive growth of data requires an automated way to extract useful knowledge. Thus, medical domain is a major area for applying data mining. Through data mining, we can extract interesting knowledge and regularities. The discovered knowledge can then be applied in the corresponding field to increase the working efficiency and improve the quality of decision making.
We developed a knowledge discovery system to extract knowledge from data. There are five steps in the system (Fig. 1). Real-life data are collected in the first step. Then, the data must be preprocessed before analysis can be started. The third and fourth step induce knowledge from the preprocessed data. The causality and structure analysis step learns the overall relationships between the variables. A resulting Bayesian network represents the knowledge structure. Based on this knowledge, the user can specify the grammar for the target rules to be discovered from data. This grammar is used for the rule learning step that learns a set of significant rules from the data. In the fifth step, the discovered knowledge is verified and evaluated by the domain experts. The domain experts may discover and correct mistakes in the discovered knowledge. On the other hand, the learned knowledge can be used to refine the existing domain knowledge. Finally, the learned Bayesian network is used to perform reasoning under uncertainty, and the induced rules are incorporated into an expert system for decision making.
In this paper, we present the two knowledge learning steps which are the core of the knowledge discovery system. They both employ evolutionary computation as the search algorithms. This paper is organized as follows. Section 2 introduces the backgrounds on evolutionary computation, Bayesian network learning, and rule learning. Section 3 describes the approaches for learning Bayesian networks. The rule learning process is delineated in Section 4 and the details of the techniques are given in Section 5. The data mining system has been applied to two real-life medical databases. The results are presented in 6 Results on the fracture database, 7 Results on the scoliosis database and the conclusion is presented in Section 8.
Section snippets
Evolutionary computation
The term evolutionary computation is used to describe algorithms that simulate the natural evolution to perform function optimization and machine learning. They are based on the Darwinian principle of evolution through natural selection. The algorithms maintain a group of individuals to explore the search space. Examples of evolutionary computation include genetic algorithms (GA) [19], [13], genetic programming (GP) [24], [25], evolutionary programming (EP) [10], [11] and evolution strategy
Causality and structure analysis
In the proposed knowledge discovery process (Fig. 1), the causality and structure analysis process induces a Bayesian Network from the data. The learning approach is based on Lam and Bacchus’s work [27], [26] on employing the minimum description length (MDL) principle to evaluate a Bayesian Network. EP is employed to optimize this metric in order to search for the best network structure.
Rule learning
The second step in our data mining process is to learn rules from the data. Our learning approach is based on generic genetic programming (GGP) [43], [42], [40], which is an extension of GP. It uses a grammar [21] to control the structures evolved in GP.
Novel techniques for rule learning
Other than using GGP as the search algorithm, other techniques are needed so as to efficiently learn multiple interesting rules from the database. These techniques are described in the following section.
Results on the fracture database
The described data mining technology has been applied to a real-life medical database consisting of children with limb fractures, admitted to the Prince of Wales Hospital of Hong Kong during the period 1984–1996. This data can provide information for the analysis of child fracture patterns. This database has 6500 records and eight attributes, which are listed in Table 3.
Results on the scoliosis database
The data mining process has been applied to the database of scoliosis patients. Scoliosis refers to the spinal deformation, where a patient suffering from this has one or several curves in his spine. Among them, the curves with severe deformations are identified as major curves. The database stores measurements on the patients, such as the number of curves, curve location, degrees and directions. It also records the maturity of the patient, the class of scoliosis and the treatment. The database
Conclusion
We have presented a data mining system that is composed of five steps. The third and fourth steps are detailed. They both employ evolutionary computation as the search algorithms. Causality and structure analysis focuses on the general causality model between the variables while rule learning captures the specific behavior between particular values of the variables.
Our system is particularly suitable to the analysis of real-life databases that cannot be described completely by just a few rules.
Acknowledgements
This work was partially supported by Hong Kong RGC CERG Grant CUHK 4161/97E and CUHK Engineering Faculty Direct Grant 2050151. The authors wish to thank Chun Sau Lau and King Sau Lee for preparing, analyzing and implementing the rule learning system for the scoliosis database.
References (43)
- et al.
Classifier systems and genetic algorithms
Artif. Intell.
(1989) The computational complexity of probabilistic inference using Bayesian belief networks
Artif. Intell.
(1990)- Agrawal R., Imielinski T., Swami A. Mining association rules between sets of items in large databases. Proceedings of...
- R.R. Bouckaert. Properties of belief networks learning algorithms. Proceedings of the Conference on Uncertainty in...
Bayesian networks without tears
AI Mag.
(1991)- et al.
Approximating discrete probability distributions with dependence trees
IEEE Trans. Inf. Theory
(1968) - et al.
The CN2 induction algorithm
Mach. Learn.
(1989) - et al.
A Bayesian method for the induction of probabilistic networks from data
Mach. Learn.
(1992) - Fayyad U.M., Piatesky-Shapiro G., Smyth P. From data mining to knowledge discovery: An overview. AI Mag.,...
An introduction to simulated evolutionary optimization
IEEE Trans. Neural Netw.
(1994)