Instance-based classifiers applied to medical databases: diagnosis and knowledge extraction.

OBJECTIVE

The aim of this paper is to study the feasibility and the performance of some classifier systems belonging to family of instance-based (IB) learning as second-opinion diagnostic tools and as tools for the knowledge extraction phase in the process of knowledge discovery in clinical databases.

MATERIALS AND METHODS

We consider three clinical databases: one relating to the differential diagnosis of erythemato-squamous diseases, the second to the diagnosis of the onset of diabetes mellitus and the third dealing with a problem of diagnostic imaging in nuclear cardiology. We apply five IB classifiers to each database; two are based on exemplars, one is based on prototypes and two are hybrid. One of the latter classifiers is a new classifier introduced here and is called prototype exemplar learning classifier (PEL-C). We use cross-validation techniques to evaluate and compare the performances of several classifier systems as diagnostic tools, considering indexes such as accuracy, sensitivity, specificity, and conciseness of class representations. Moreover we analyze the number and the type of instances that represent the diagnostic classes learnt by each classifier to evaluate and compare their knowledge extraction capabilities.

RESULTS

An examination of the experimental results shows that classifiers with the best classification performances are the optimized k-nearest neighbour classifier (k-NNC) and PEL-C. The k-NNC uses the highest number of representative instances, 100% of the entire database, whereas PEL-C uses a far lesser number of representative instances: equal, on the average, to the 3% of the database. As tools for knowledge extraction, we interpret the kind of class representations obtained by IB classifiers as a form of nosological knowledge. Additionally, we report the most interesting diagnostic class representations to be those extracted by PEL-C because they are composed of a mixture of abstracted prototypical cases (syndromes) and selected atypical clinical cases.

CONCLUSION

This study shows that IB methods - most notably, the optimized k-NNC and the PEL-C - can be used and may be advantageous for clinical decision support systems and that IB classifiers can be used for nosological knowledge extraction. Because PEL-C uses more compact and potentially meaningful class descriptions, it is preferable when the diagnostic problem at-hand needs smaller storage space or for knowledge extraction itself. The complexity and responsibility of diagnostic practice requires that these results be confirmed further within other clinical domains.