Knowledge mined from clinical data can be used for medical diagnosis and prognosis. By improving the quality of knowledge base, the efficiency of prediction of a knowledge-based system can be enhanced. Designing accurate and precise clinical decision support systems, which use the mined knowledge, is still a broad area of research. This work analyses the variation in classification accuracy for such knowledge-based systems using different rule lists. The purpose of this work is not to improve the prediction accuracy of a decision support system, but analyze the factors that influence the efficiency and design of the knowledge base in a rule-based decision support system. Three benchmark medical datasets are used. Rules are extracted using a supervised machine learning algorithm (PART). Each rule in the ruleset is validated using nine frequently used rule interestingness measures. After calculating the measure values, the rule lists are used for performance evaluation. Experimental results show variation in classification accuracy for different rule lists. Confidence and Laplace measures yield relatively superior accuracy: 81.188% for heart disease dataset and 78.255% for diabetes dataset. The accuracy of the knowledge-based prediction system is predominantly dependent on the organization of the ruleset. Rule length needs to be considered when deciding the rule ordering. Subset of a rule, or combination of rule elements, may form new rules and sometimes be a member of the rule list. Redundant rules should be eliminated. Prior knowledge about the domain will enable knowledge engineers to design a better knowledge base.