Knowledge-based data analysis and interpretation

Perhaps the most distinctive feature that accompanies medical data analysis is knowledge; data analysis in medicine strives to discover new and useful knowledge, while using available knowledge to guide the process and incorporate it into discovered models. In these terms, we perform data analysis to extract new findings that either refine or supplement existing knowledge on the problem domain, a population of patients, or a specific patient under consideration. We see this interaction between knowledge and analysis of medical data and interpretation of its results in a knowledge management perspective [1] as a knowledge cycle (Fig. 1). Domain knowledge influences what kind of experiments we will perform , what phenomenon we will observe, what kind of data will we gather, or what data base or repository we will query. The data we obtain in this way is analyzed and results are interpreted to test our working hypothesis, potentially discovering new findings and verifying their consistency with the existing body of knowledge, which is updated accordingly. In data analysis, knowledge is represented either implicitly or explicitly. By explicit knowledge we refer to knowledge already established, formalized and coded within some knowledge base. Such knowledge is used in some way in the process of data analysis. Researchers in machine learning often refer to such knowledge as ''background knowledge ,'' and use it in learning either in data prepro-cessing, feature selection or modeling. While this approach seems very promising in medical applications , the number of methods, tools and applications of such an approach are few. Most applications of data mining in medicine, for instance, have focused on building models directly from data and do not consider any explicitly represented knowledge in the process. These applications leave the interpretation of the models and the placement of any new information found in this way within the context of the available knowledge to analysts and domain experts. Present approaches of data analysis in medicine more often rely on existing knowledge implicitly; it is important to note that such knowledge is not restricted to the clinical or scientific domain, but also includes analytic knowledge. We mention just four examples here: The process of feature selection, or determining the set of variables that are thought best, a priori, to contribute to an analysis depends on an intimate familiarity with the problem domain. Numerous methods are available for reducing the feature space. Some are cognitive, such as expert …